Constraint-Based Clustering to Find Communities in a ...

Proceedings

Information Science and Industrial Applications

ISI 2012 May 29-31, 2012 Waterfront Airport Hotel and Casino, Cebu, Philippines

Volume Editors: Chin-Chen Chang Feng Chia University, Taiwan Email: [email protected] Yvette E. Gelogo Hannam University, Korea Email: [email protected] Ronnie D. Caytiles Hannam University, Korea Email: [email protected]

Copyright  Science and Engineering Research Support soCiety (SERSC) All rights reserved

Copyright and Reprint Permissions: Abstracting is permitted with the credit to the source. Other copying, reprint request should be addressed to: SERSC Copyrights Manager, Head Office: 20 Virginia Court, Sandy Bay, Tasmania, Australia, Phone no.: +61 3 9016 9027, Fax no.: +82-42-367-0859

ISSN 2287-1233

Foreword

Information Science and Industrial Applications are areas that attracted many academic and industry professionals to research and develop. The goal of this conference is to bring together the researchers from academia and industry as well as practitioners to share ideas, problems and solutions relating to the multifaceted aspects of Information Science and Industrial Applications. We would like to express our gratitude to all of the authors of submitted papers and to all attendees, for their contributions and participation. We believe in the need for continuing this undertaking in the future. We acknowledge the great effort of all the Chairs and the members of Editorial Committee of the above-listed event. Special thanks go to SERSC (Science & Engineering Research Support soCiety) for supporting this conference. We are grateful in particular to the following speaker who kindly accepted our invitation and, in this way, helped to meet the objectives of the conference: Prof. Chin-Chen Chang, Feng Chia University, Taiwan.

May 2012

Chairs of ISI 2012

Preface

We would like to welcome you to the Regular Papers proceedings of Information Science and Industrial Applications (ISI 2012) which was held on May 29-31, 2012, at Waterfront Airport Hotel and Casino, Cebu, Philippines. ISI 2012 is focused on various aspects of advances in Information Science and Industrial Applications. It provided a chance for academic and industry professionals to discuss recent progress in the related areas. We expect that the conference and its publications will be a trigger for further related research and technology improvements in this important subject. We would like to acknowledge the great effort of all the Chairs and members of the Editorial Committee. We would like to express our gratitude to all of the authors of submitted papers and to all attendees, for their contributions and participation. We believe in the need for continuing this undertaking in the future.

May 2012 Chin-Chen Chang Ronnie D. Caytiles Yvette E. Gelogo

Organization

General Chair Ruay-Shiung Chang, National Dong Hwa University, Taiwan Program Chair Javier Garcia-Villalba, Universidad Complutense of Madrid, Spain Editorial Committee Jin Wang, Nanjing University of Information Science & Technology, China Hsi-Ya Chang (Jerry), National Center for High Performance Computing, Taiwan J. H. Abawajy, Deakin University, Australia Sabah Mohammed, Lakehead University, Canada Wai-chi Fang, National Chiao Tung University, Taiwan Adrian Stoica, NASA JPL, USA Aboul Ella Hassanien, Cairo University, Egypt Dominik Slezak, Warsaw University & Infobright, Poland Gongzhu Hu, Central Michigan University, U.S.A Kirk P. Arnett, Mississippi State University, U.S.A Sankar Kumar Pal, Indian Statistical Institute, India Martin Drahansky, BUT, Faculty of Information Technology, Czech Republic Filip Orsag, BUT, Faculty of Information Technology, Czech Republic Samir Kumar Bandyopadhyay, University of Calcutta, Kolkata, India Tadashi Dohi, Hiroshima University, Japan Tatsuya Akutsu, Kyoto University, Japan Carlos Ramos, GECAD and ISEP, Portugal Muhammad Khurram Khan, King Saud University, KSA Hideo KURODA, FPT University, Vietnam Wenbin Jiang, Huazhong University of Science & Technology, china Tao Gong, Donghua University, China Haengkon Kim, Catholic University of Daegu, Korea Byeongho Kang, University of Tasmania, Australia Jin Wang, Nanjing University of Information Science & Technology, China

Table of Contents

Constraint-Based Clustering To Find Communities In A Social Network ............ 1 Soumita Seth, Debnath Bhattacharyya, Tai-hoon Kim Secure Data Hiding in Binary Text Document for Authentication ...................... 10 Asmita Haveliya, Debnath Bhattacharyya, Tai-hoon Kim Probability Inclusion Degrees and Probability Fuzzy Attribute Implications for Interval-Valued Fuzzy Sets ............................................................................ 21 Yanhui Zhai, Deyu Li, Kaishe Qu Analytical Evaluation of MDPSK and MPSK Modulation Techniques over Nakagami Fading Channels .................................................................................. 27 Alam S. M. Shamsul, Kwon GooRak, and Choi GoangSeog A Fuzzy Clustering Approach for Missing Value Imputation with Non-Parameter Outlier Test .......................................................................... 33 Jing Tian, Bing Yu, Dan Yu, Shilong Ma Web Robot Detection Based on Monotonous Behavior ...................................... 43 Shinil Kwon, Myeongjin Oh, Dukyun Kim, Junsup Lee, Young-Gab Kim, and Sungdeok Cha A Configuration Management Model on the High-speed Networks ................... 49 Jie Huang, Lin Chen Challenges and Solutions of Cyber-Physical Systems ......................................... 55 Lichen Zhang, Jifeng He and Wensheng Yu Red Tide Image Recognition Using Roundness and Entropy Based on Semantic Feature ............................................................................................. 61 Sun Park, Yoo Kang Ji, ByungRae Cha A Study of Zigbee Network Topology Organization for Reliable Data Transmission ......................................................................................................... 67 Sung-Je Woo

Pattern Mining and Querying in Business Intelligence ........................................ 73 Nittaya Kerdprasop, Fonthip Koongaew, Phaichayon Kongchai, Kittisak Kerdprasop An Improved Low Complexity MIMO-OFDM Detection Algorithm ................. 79 Hye-Yeon Jeong, Jang-Kyun Ahn, Hyun-Woo Jang and Hyoung-Kyu Song A Cooperative Communication Technique for Improved Transmission Rate in OFDMA-Based System ....................................................................................... 84 Hyoung-Muk Lim, Eui-Hak Lee, Hyun-Woo Jang and Hyoung-Kyu Song Efficient Wireless Communication System for Industrial Communications ....... 89 Jung-In Baik, Hye-Yeon Jeong, Jang-Kyun Ahn and Hyoung-Kyu Song Link Adaptive Cooperative Communication with Smart Relays for MIMOOFDMA ............................................................................................................... 94 Jung-In Baik, Eui-Hak Lee, Hye-Yeon Jeong and Hyoung-Kyu Song Implementation of a Face Recognition System for Interactive TV Control System................................................................................................................... 98 Sang-Heon Lee, Myoung-Kyu Sohn, Dong-Ju Kim, Byungmin Kim, Hyunduk Kim, and Chul-Ho Won An Energy-Efficient Dynamic Layered Coverage Deployment Mechanism in Wireless Sensor Network.................................................................................... 104 Bing Yu, Jing Tian, Dan Yu, and Shilong Ma Adaptive Spectrum Sensing Algorithm for Cognitive Radio System ................ 113 Mi Sun Lee, Yoon Hyun Kim, Jin Young Kim Performance of Efficient Signal Detection Scheme for Heterogeneous Network Structure ............................................................................................... 119 Kyung Sun Lee, Yoon Hyun, Kim, Jin Young Kim Analyzing and Optimizing the Combined Primality Test with GCD Operation On Smart Mobile Devices................................................................................... 127 Hosung Jo and Heejin Park

A visible obstacle mobility model based on activity area ................................... 133 Zhou Shasha, Wang Dong, Li Xiaohong A New Algorithm for Finger-Vein Image Enhancement and Segmentation ..... 139 Yihua Shi, Jinfeng Yang and Jucheng Yang An Architecture Model of Sensor Information System Based on Cloud Computing .......................................................................................................... 145 Pengfei You, Yuxing Peng Analysis of the model DEA-CCR in the Selection of Investments of the Pension Funds Entities.............................................................................. 151 Átila Einstein de Oliveira, Plácido Rogério Pinheiro Application Hybrid Model to Support in the Diagnosis of Posttraumatic Stress Disorder .................................................................................................... 157 Luciano Comin Nunes, Plácido Rogério Pinheiro, Tarcísio Cavalcante Pequeno, Mirian Calíope Dantas Pinheiro Image Latent Semantic Analysis for Face Recognition ...................................... 164 Jucheng Yang Impact on Applying Cloud Computing Service to IT Education........................ 170 Chen-Feng Wu Dynamic Resource Management for Energy Saving in the Cloud Computing Environment ....................................................................................................... 176 Liang-Teh Lee, Kang-Yuan Liu, and Hui-Yang Huang A Systematical Digital Content Management Based on the Unification of Its Process and the Standardization of Content Format ........................................... 182 Wongoo Lee, Minho Lee, Yunsoo Choi, Donghoon Choi, Soonchan Hong, Sakwang Song, Hanmin Jung, Hwamook Yoon An Approach to the Verification and Analysis of Cyber Physical Systems ....... 188 Wei Qiu, Hua-qing Min, Li-chen Zhang

Access Control Mechanism supporting Scalability, Interoperability and Flexibility of Multi-Domain Smart Grid System ......................................... 194 Jihyun Kim, Yanghyeon Kwon, Yunjin Lee, Jungtaek Seo, and Howon Kim A Study on Data Communication Method for Optimizing Communication Cost of Real-Time Based Global Container Tracking System ........................... 202 Joong Jo Shin, Hyung Rim Choi, Kang Bae Lee, Jung Rok Son, Eun Kyu Lee, Jin Wook Lee, Hee Mok Son A Modified DRR-Based Non-real-time Service Scheduling Scheme in Wireless Metropolitan Networks .................................................................... 209 Han-Sheng Chuang, Liang-Teh Lee and Chen-Feng Wu A Novel Image Preprocessing Technique Based on Binary Patterns For Furthermore Processing................................................................................ 215 Dong-Ju Kim, Sang-Heon Lee, Myoung-Kyu Sohn, Byungmin Kim, Hyunduk Kim An Ontology Based Context Aware System to Control the Robot Manipulator in Delivery Logistics Industry ............................................................................ 221 Taeho Kim, Dongmun Shin, Myungsic Kim, Jinsuk Kim, Keunho Ryu, Moonsun Shin Security Engineering Approach for the Development of Secure Information Systems ............................................................................................................... 227 Young-Gab Kim and Sungdeok Cha A New Framework of Music Recommendation Using Collaborative Filtering............................................................................................................... 233 Hyun-Tae Kim, Chang Wook Ahn Performance Analysis of an FMS with Finite Capacity Buffers and Single AGV ................................................................................................. 239 Soo Tae Kwon Integer Linear Programming for Detecting Hedges in Natural Language Text ................................................................................................................... 246 Marcelo Coelho Quinho, Cicero Nogueira dos Santos and Placido Rogerio Pinheiro

Applying the Generate and Solve Methodology in the Problem of Dynamic Coverage and Connectivity in Wireless Sensor Networks ................................. 252 Placido R. Pinheiro, Andre L. V. Coelho, Alexei B. Aguiar, Alvaro M. S. Neto Anatomy of Biometric Passports ........................................................................ 258 Dominik Mal and Martin Drahansk´y A Novel Adaptive Virtual Machine Deployment Algorithm for Cloud Computing .......................................................................................... 264 Hongjae Kim, Munyoung Kang, Sanggil Kang, Sangyoon Oh The Design of a Web Contents Management System......................................... 270 Jaegeol Yim , Gyeyoung Lee The Functional User Requirement Analysis of a Web Broadcasting Management System ........................................................................................... 276 Gyeyoung Lee, Jaegeol Yim Lane Detection using Fuzzy C-Means Clustering .............................................. 282 Kwang-Baek Kim, Doo Heon Song, Jae-Hyun Cho Performance Evaluation of HEVC Intra Prediction Algorithms Parallelized in a Mesh-based Multicore System .................................................................... 288 Neungsoo Park, Young-ho Park and Yungho Choi A Proactive Anti-Phishing Tool Using Fuzzy Logic and RIPPER Data Mining Classification Algorithm ........................................................................ 292 Rosana J. Ferolin Real Time Inverse Kinematics Robot Gait Sequencing Using Self Organizing Maps................................................................................................. 305 Chris Jordan G. Aliac Agent-based Safety Support System for Collision Avoidance in Dilemma Zones................................................................................................................... 314 Eunyoung Lee, Hyung-Ju Cho, Kiyeol Ryu

Sensor Data Network Using Android Mobile Devices ....................................... 324 Chris Jordan G. Aliac Ontology-based RFID Code Conversion for Ubiquitous Applications .............. 334 Young-Sik Noh, Yung-Cheol Byun, SangJoon Lee Location Recognition of Wireless Sensor Nodes Based on Environmental Data Classification .............................................................................................. 338 Ae-cheoun Eun and Young-guk Ha

Constraint-Based Clustering to Find Communities in a Social Network Soumita Seth1, Debnath Bhattacharyya2, Tai-hoon Kim3. 1

Vinod Gupta School of Management, IIT Kharagpur, Kharagpur, India. [email protected] 2 Department of Information Technology, MCKV Institute of Engineering,Liluah, Howrah711204, India. [email protected] 3 School of Information Science, Global Vision School Australia (GVSA), 20 Virgina Court, Sandy Bay, Tasmania, Australia [email protected]

Abstract. Social network analysis is used to improve the performance of userspecific information dissemination in many internet-based applications like, social networking websites, on-line viral marketing, and recommendation network based applications. Detecting communities, which are basically subgraphs or clusters, within a social network has been the central focus of this work. Here, we present a divisive hierarchical clustering procedure based on minimum edge-cut principle that is already followed by CHAMELEON: Two Phase Agglomerative Hierarchical Clustering [1]. As the stopping criteria of this algorithm depends on two threshold constraints namely, balance constraint(BC) and MINSIZE (MS), it is termed as Constraint-Based Clustering .As per our claim this divisive hierarchical clustering method divides the nodes of social network into disjoint communities by removing minimum number of edges. As a measure of the quality of community, we follow network centrality measure clustering coefficient. Our experimental results, using some well-known benchmark social networks, also show that our method determines similar communities with good average clustering coefficient as the other existing well known methods of various research papers. Keywords: Social network analysis, Community detection, Hierarchical clustering, Minimum edge-cut principle, Clustering coefficient.

1

Introduction

NETWORKS are ubiquitous in the scientific models around us. Social networks consist of individuals, who are represented as nodes and relationships between them are represented by the edges between the nodes. In recent years, growth of web-based communication has triggered an enormous amount of social interactions among the users on internet. This data can generate a large social network with millions or even billions of interactions. Such social networks are difficult to handle because of its large size but consist of interesting inherent properties like presence of social communities and the structures of those communities. A graph can be defined by G =

1

(V, E), where V= {vi : 1 ≤ i ≤ n} is the set of nodes and E is the set of edges or links. We write n = |V|, m = |E|, eij = (vi , vj ) = (vj , vi ) for the undirected edge connecting vi and vj (≠ vi), Nghb(vi)= {vj : eij E } and di = degree (vi) = |Nghb(vi)| [2]. In data mining clustering is a process which can groups a set of data in a way such that it can maximize the intra-cluster similarity and minimize the inter-cluster similarity [1]. Among the existing clustering algorithms K-means, PAM [3, 1], and CLARANS [8, 1] assume that clusters are hyper-ellipsoidal (or globular) and are of similar sizes. DBSCAN [3, 1] assumes that all points within genuine clusters are density reachable and points across different clusters are not. Agglomerative hierarchical clustering algorithms CURE and ROCK [3, 1] use a static model to determine the most similar cluster to merge. CURE measures the similarity of two clusters based on the similarity of the closest pair of the representative points belonging to different clusters, without considering the internal closeness (i.e., density or homogeneity) of the two clusters involved. ROCK measures the similarity of two clusters by comparing the aggregate inter-connectivity of two clusters against a userspecified static inter-connectivity model, and thus ignores the potential variations in the interconnectivity of different clusters within the same data set. These algorithms can breakdown if the choice of parameters in the static model is incorrect with respect to the data set being clustered, or if the model is not adequate to capture the characteristics of clusters. Furthermore, most of these algorithms breakdown when the data consists of clusters that are of diverse shapes, densities, and sizes. Besides above described clustering algorithms, one existing renowned hierarchical clustering algorithm is CHAMELEON [1, 3] that measures the similarity of two clusters based on a dynamic model. Unlike CURE and ROCK CHAMELEON consider the internal inter-connectivity of the clusters and closeness of items within the clusters and if both are comparable then two clusters are merged. CHAMELEON is efficient in a number of data sets that contains points in 2D space, and contain clusters of different shapes, densities, sizes, noise and artifacts [1]. CHAMELEON represents the set of data items as a sparse graph based on knearest neighbour graph approach [3, 1] to scale large data sets and operate on dataset successfully that are available in similarity space and not in metric spaces. CHAMELEON clustering algorithm is consist of two phases: Finding Initial Sub Clusters and Merging Sub-Clusters using a Dynamic Framework. (refer Fig.1).

Fig 1. Overall framework CHAMELEON, (source: [3, 1]).

In this paper we present a Divisive Hierarchical Algorithm whose idea is based on the partition technique of first phase of CHAMELEON [3, 1]. For finding the initial sub clusters CHAMELEON uses a graph partitioning algorithm which is part of hMETiS library to partition the k-nearest neighbour graph of the data set into a large number of partitions in such a way by which the edge cut i.e. the sum of the weight of the edges whose removal give partitions, is minimized. As the each edge of k-nearest

2

neighbour graph represents the similarity among data points, so minimizing edge cut implies effectively minimizes the relationship among data points across the resulting partitions. So, the data points which reside in same partition are highly related to each other. CHAMELEON initially starts with all the points in same cluster Ci and bisects it using hMETiS as example two sub clusters CiA and CiB such that edge cut between CiA and CiB is minimized and repeatedly selects the largest sub-cluster among the current set of sub-clusters to bisect if local threshold balance constraint (BC) is satisfied i.e. each of these two sub-clusters have to be consisted of at least 25% of the nodes which are present in cluster Ci. This process terminates when the larger subcluster contains less than a specified number of vertices, termed as MINSIZE. MINSIZE generally controls the granularity of the initial clustering solution. Generally setting MINSIZE to about 1% to 5% of the overall data points worked well enough. CHAMELEON assumes that after sparsification and graph partitioning produced partitions are in most cases sub-clusters. That means, most of the points in a partition belong to a true cluster [1]. The rest of the paper is organized as follows. Section 2 describes primary goal of this paper. We present our new clustering algorithm in Section 3. Section 4 gives the experimental results. Section 5 contains conclusions and directions for future work.

2

Primary Goal

After selecting social networks with prior knowledge about the groups our primary goal is to detect clusters representing communities from those social networks which are represented as undirected and unweighted graph G = (V, E) where V is the number of vertices and E is the number of edges. Like Chameleon, We have used Minimum Edge Cut Principle in our algorithm. Using adjacency-list representation of graph, i.e. adjList[G] we can find the neighbours or adjacent vertices for each vertex v ε V(G). From the neighbours list of each vertex we have found the list of common neighbours between a pair of vertices. Then we have cut or removed the edges between such pair of vertices which have no or lesser number of common neighbours. In this way we will be able to find out clusters by removing minimum number of edges. After removing edges we have used depth first search traversal operation [4] to find out a connected component of graph G. Our aim is to keep every vertex in such group where it has maximum number of neighbours. This means each partition or cluster will have to be nothing but a dense sub-graph and hence produce good quality clusters. In our algorithm there is also concept of two threshold constraints balance constraint (BC) and MINSIZE (MS) whose computation process is same with CHAMELEON clustering algorithm. In our algorithm, we have varied both local and global constraints and observed the result in each time. In this paper we have proposed a recursive hierarchical divisive clustering algorithm to detect clusters and used notion of average clustering coefficient to measure the correctness of clusters. We will discuss our algorithm in next section. We need to calculate average local clustering coefficient for two sub graphs. Clustering Coefficient Clustering coefficient is nothing but a measure of degree to which nodes in a graph tend to cluster [6].

3

Local Clustering Coefficient: Local Clustering Coefficient calculates the probabilities of number of adjacent nodes are connected with a single node. It is calculated by the fraction of permitted edges between the neighbours of a single node to them that actually exist for these neighbours [7]. (1)

Ci = According to Watts and Strogatz, the clustering coefficient for whole network is defined by the average of local clustering coefficients of all the vertices n: [6] Network Average Clustering Coefficient =

(2)

In our experiment, we have used notion of network average clustering coefficient as quality function to measure the goodness of clusters on basis of benchmark.

3

Communities in a Social Network

Communities, in a social network, are generally defined as a group of nodes that are densely connected. The density in some cases has been defined by degrees [9] and in some cases by its cliquishness, i.e., its average clustering coefficient [6]. As cliquishness denotes the true coherence between all the nodes of a group, we choose clustering coefficient as our measure to determine the quality of a community. Here we will discuss our algorithm. Definition of important terms MINSIZE (MS) ← Minimum percentages of the total number of vertices of graph G have to be present in clusters. adjM(G) ← Adjacency matrix of graph G NCM (G) ← Numbers of neighbours for each vertex v VG adjL (G) ← Adjacency list of graph G CNM (G) ← Number of common neighbours for a pair of vertices u, v u,v VG min ← Minimum number of neighbours between a pair of vertices. resultM(G) ← Result of a mathematical operation by adjM(G) and CNM(G) Balance Constraints (BC) ← Minimum percentage of vertices in a specific cluster of G has to be present in sub-clusters NgBv ← Set of neighbours for each vertex v, DEL ← Set of distinct elements of resultM (G) which are greater or equal 1 CL1 (G) ← Set of vertices in first partition/cluster of Graph. CL2 (G) ← Set of vertices in second partition/cluster of Graph. DFSSourceSelection ← Select a vertex by which we can get maximal connected component through DFS. DFSRecursive (adjL, Source, V) ← Performs DFS

4

5

In our algorithm our goal is getting several clusters by following Minimum edge cut principal. For this purpose here we always try to take care about how many number of edges essential to remove to get proper clusters. Our objective is this number of edges should be minimum. To achieve our goal here we used an objective function which is defined at Line 5 of our Algorithm 1. Here, )= ) = (4) So, we can conclude that dij represents how connected vi and vj are to their common neighbours. di and dj represent the degree of vi and vj which is nothing but and respectively. So in mathematical notation our main objective function is

=

(5)

When resultij gives minimum value that implies that we can split vi and vj into two clusters by removing minimum number of edges i.e. through sparse region. So for removing edges if we check this objective function in every iteration, we may achieve to our goal. Let’s take an example; In Fig. 3, there is a connected graph G which has a

Fig. 3. Connected Graph with set of vertices. set of vertices V (G) = {A0,A1,A2,A3,A4,B0,B1,B2,B3,B4}. Now we will apply our = = objective function on this graph. Here, = 0.75. This value is same for (A0, A2), (A0, A3), (A0, A4), (A1, A2), (A1, A3), (A1, A4), (A2, A3), (A2, A3), (A3, A4) as this sub graph is a completely connected graph. Now = = =0 You may check rest of vertices B0, B1, B2, B3, B4 in same way but here minimum value is 0. So if we remove the edge e (A0, B0), we will get two connected components which are separated by sparse region. For checking connectivity and cycle of the graph we used Depth First Search (DFS) operation [4].

4

Experimental result

4.1 Experimental Setup We have written a C program for the experimental study, and executed it on a Compaq 510 laptop with 2.00 GHz Intel(R) Core(TM)2 Duo CPU T8570 microprocessor and 2 GB RAM memory, running Fedora (2.6.23.1-42.fc8). One more experiment has been performed for computing network average clustering coefficient, using cluster sets extracted from the various social network databases. In our experiment, we have used three different social networks for detecting community from each network. Our first social network is Zachary’s Karate Club Members group (Node: 34, Edge: 78), Dolphin Social Network (Node: 62, Edge: 159) and American Football Network (Node: 115, Edge: 613). From these social networks

6

we have found different set of clusters based on different balance constraint (BC) and MINSIZE (MS) which is said in CHAMELEON’s first phase. As the determination of values of BC and MS is completely depends on nature of graph, for different graphs we may choose different set of thresholds by trial and error method.

4.2 Result for Different Data sets Zachary Karate Club Data: This is a social network of 34 karate club members in an US university in 1970s. W. Zachary observed and recorded the interactions between individuals for 2 years and this social network data has been used for benchmarking in many papers related to community detection. The club got divided because of a conflict between administrator and instructor. In Fig. 4, we show a visualization of Karate Data. For this network we have taken several combination of thresholds which are {BC 15% MS 3% or 5%}, {BC 25% MS 3% or 5%} to computed average clustering coefficient for every clusters record and in Fig. 5, we got maximum average clustering coefficients 0.686 when number of clusters 2 by all combinations. Here we also compare average clustering coefficient of our clusters with the clusters record of well known benchmark algorithm Strength of Link (SOL). It’s clear that our algorithm gives similar clusters of SOL.

Fig. 4.

Visualization o f Karate Club Data

Fig. 5. Average Clustering Coefficient Plot of Zachary Karate Club Data and Compare with Benchmark Clustering Algorithm Strength of Link (SOL).

Dolphins Social Network: For this network we have taken several combination of thresholds which are {BC 10% MS 3%}, {BC 15% MS 3%}, {BC 25% MS 3%}

7

to computed average clustering coefficient for every clusters record and in Fig. 6, we got maximum average clustering coefficients 0.566, 0.589, 0.938, 0.589 when number of clusters 8, 6, 5 respectively. Here we also compare average clustering coefficient of our clusters with the clusters record of well known benchmark algorithm Strength of Link (SOL). It’s clear that our algorithm gives similar clusters of SOL. American Football Network: For this network we have taken several combination of thresholds which are {BC 5% MS 3%}, {BC 10% MS 3%}, {BC 5% MS 5%}, {BC 10% MS 5%} to computed average clustering coefficient for every clusters record and in Fig. 7, we got maximum average clustering coefficients 0.943, 0.943, 0.938, 0.911 when number of clusters 15, 15, 14, 13 respectively. Here we also compare average clustering coefficient of our clusters with the clusters record of well known benchmark algorithms Girvan-Newman (GN) which is based on Edge Betweeness and Strength of Link (SOL). It’s clear that our algorithm gives better clusters than GN and similar clusters of SOL.

Fig. 6. Average Clustering Coefficient Plot of Dolphins Social Network and Compare with Benchmark Clustering Algorithm Strength of Link (SOL).

Fig. 7. Average Clustering Coefficient Plot of American Football Network and Compare with Benchmark Clustering Algorithms Girvan Newman (GN) and Strength of Link (SOL).

8

5

Conclusion and future scope

In this paper, we have presented a divisive hierarchical clustering algorithm based on minimum edge-cut principle to detect disjoint communities in a network. We have used average clustering coefficient as a quality function to measure the quality of clusters. The experiment shows that for every network which we have used in our experiment, like, Zachary’s Karate Club Member Network, Dolphins Social Network, American Football Network which have inherent communities, our algorithm can detect communities and give meaningful clusters that are similar to the clusters generated by other sophisticated graph clustering methods. In our algorithm our objective is detect communities from the network by cutting minimum number of edges and from our experimental result we can say that our criteria has fulfilled, we have got dense region of graphs as clusters divided by sparse region. For the extension of this paper we will measure our cluster’s quality by the notion of modularity as quality function and compare the result with results of benchmark in some research papers. Beside this, in future we want to include the concept of overlapping clusters in this algorithm. An interesting extension of this work would be to test our algorithm on large datasets that can be represented as weighted graphs.

References 1.

George Karypis, Eui-Hong(Sam) Han and Vipin Kumar, “CHAMELEON: A Hierarchical Clustering Algorithm Using Dynamic Modeling,” Technical Report #99007, Department of Computer Science and Engineering, University of Minnesota, Minneapolis, MN 55455, USA. 2. R. S. Salaria, Data Structures & Algorithms Using C, Khanna Book Publishing Co. (P) LTD., New Delhi, 2004. 3. Pang-Ning Tan, Vipin Kumar and Michael Steinbach, Introduction To Data Mining, Prearson Education, Inc. , New Delhi, 2006. 4. Thomas H. Cormen, Charles E. leiserson, Ronald L. Rivest and Clifford Stein, Introduction To Algorithms, Prentice-Hall India (PHI), New Delhi, 2001 5. Jiaweli Han and Micheline Kamber Data Mining: Concepts and Techniques, Second Edition, Morgan Kaufmanu Publishers, March 2006 6. http://en.wikipedia.org/ wiki/Clustering_coefficient 7. http://sonivis.org/ wiki/ index.php/Local_Clustering _Coefficient 8. Erica Kolatch “Clustering Algorithms for Spatial Databases: A Survey,” Technical Report CMSC 725, Department of Computer Science, University of Maryland, College Park, March 2001. 9. Vinayaka Pandit, Natwar Modani,Sougata Mukherjea, Amit Anil Nanavati, Sambuddha Roy, Amit Agarwal “Extracting dense communities from telecom call graphs”, COMSWARE 2008:82-89. 10. Jure Leskovec, Ajit Singh, Jon M. Kleinberg “Patterns of Influence in a Recommendation Network”, PAKDD 2006: Pages 380-389.

9

Secure Data Hiding in Binary text Document for Authentication Asmita Haveliya1, Debnath Bhattacharyya2, Tai-hoon Kim3. 1

Department of Electronics, ASET, Amity University, Lucknow, India. [email protected] 2 Department of Information Technology, MCKV Institute of Engineering, Liluah, Howrah-711204, India. [email protected] 3 School of Information Science, Global Vision School Australia (GVSA), 20 Virgina Court, Sandy Bay, Tasmania, Australia. [email protected]

Abstract: In an effort to detect a secure technique for the text document, we examined the Steganography techniques for secure data transmission. Steganography deals with the understanding of information hiding in a way that only sender and receiver can understand and realize the subsistence. It is a method of hiding the transmitted data in the form of codes that is hiding the data in image text etc, although in the proposed work we are concerned about the text data hiding in text document, and thus the author of the proposed work has used illustrations in DOC, PPT, TXT, MATLAB script file(.M) formats. There are various methods of information hiding but in text Steganography it is not easy to hide information as text data provides us less redundant space for concealing the secrets. Simulation outcomes demonstrate that data embedding based on coding procedure pursued has no traceable distortions in the host file. The results for the proposed work are obtained using the MATLAB version 7.10.0.499(R2010a). Keywords: Steganography, hiding data, text file encryption, text file decryption, covert writing, ppt cover data, doc Steganography.TXT and MATLAB SCRIPT files cover data

1 Introduction Before having a discussion on the subject of Steganography, let’s deal with information hiding. Information hiding consists of processes and schemes to bury or cover the information from intermediary, for instance hackers. The sender requirements is to put out of sight his data from third parties and also clearly has desires to communicate it through a technique with the intention of no more than receiver obtains it. For this purpose we require a technique and Steganography is a tool which provides the accessibility to the sender to hide from view the transmitted information. The secret message file and the cover file in which the message file is to be embedded can be text, any image, video, an audio, executable file and so forth.

10

Sending information clandestinely and communicating covertly encompass an immense amount of attention for epoch. These data hiding technique is exceptionally to a large extent in use from ancient times. It was as well employed by German Nazis soldiers to dispatch the messages. Invisible ink was also one method of Steganography for information hiding in ancient epoch. for the duration of World War II, undetectable invisible inks presented a common form of invisible writing. With the invisible ink, a seemingly innocent letter could contain a very different message written between the lines. Therefore, the document text can conceal a hidden meaning through the use of unfounded ciphers (unencrypted message), which absolute perfectly camouflage the genuine real message in an ordinarily appearing letter. There are quite a lot of methods [1-9] for this and even more highly secured methods are under investigation and are increasing gradually. If we have a discussion on the subject of text Steganography, then we will become conscious of the fact that the chief concern in text Steganography is the redundancy of dataBrassil [8] and Bender [7] makes an expression that text Steganography is the most difficult kind of Steganography that was due to redundancies. During recent times, in the field of text Steganography it was specifically designed and deliberated to take advantage of the specific, detailed characteristic of the target language. Here, in the proposed work author deals not with the specific, detailed characteristic of any targeted language, but with the file types that can have text information inside like the word file with .DOC extension, PowerPoint file with .PPT extension, the notepad file with .TXT extension, and the MATLAB script file with .M extension. This method gives the advantage of having any language involved; it is a upgrading in the enhancement of the Steganography technique which now can be applied to a file type, which is capable to encompass whichever language. This facilitates programmer not to have knowledge of all the languages actually being worked on.

2 Literature Survey Mohammad Shirali-Shahreza in his paper introduces the method which uses, changing the word of American and England language to hide the information. There are many words in both languages which are pronounced same but are different in spelling and this technique deals with them to give privacy to information which has to be transmitted. We can use either words or spelling for hiding, for example: “dialog” is an American word which is “dialogue” in British language and from these types of words having similar sound but different in character an efficient way of information hiding takes place which is called as text Steganography[1]. M. Hassan Shirali-Shahreza and Mohammad Shirali-Shahreza, [2] and A. Gutub and M. Fattani [9] in their paper proposed to hide the information in Persian and Arabic words. Arabian and Persian are very much similar language and only have

11

four different letters. These two languages collapse to generate the message which hides the data bits. Well, L. Y. POR, B. Delina in 2008 introduced approaches of text Steganography which is a part of information hiding. Text Steganography is the art of sending message hidden into texts. This document deals with inter word and inter paragraph spacing to hide the data which generate a cover page dynamically[3]. Nuur Alifah Roslan, Ramlan Mahmod, Nur Izura Udzir in their documentation deals with the text Steganography in Arabic text. Arabic text Steganography uses a sharp edges method, which allows one to hide the secret bits in the sharp-edges for each character in the Arabic text document. This technology is very much dedicated to the hiding capacity of information. This technique allows hiding the information more in amount and hence very much needed that it is introduced to public also [4]. Srivastav et al proposed three new ways for Hindi text Steganography which is based on the use of directory/dictionary structure.forst method is Punctuation Marks Based, second is Synonym Based while third is Sanskrit Classification Based [5]. Banarjee et al also introduced a new code representation technique (SSCE - Secret Steganography Code for Embedding) at both ends of communication channel. Before the embedding operation each character of the secret message has been converted to SSCE Value and then embeds to cover text and at decoding end the process reverse [6].

3 Proposed Work The process of data hiding through Steganography in the proposed work pursue the following steps, first of all it reads the input cover file and the secret file that is to be embed subsequently it checks if the size of the secret file is less than the 5% of the size of the cover file, it will be encrypted. Otherwise an error message is displayed and will ask to replace the existing cover file with whichever file of large size, then user will necessitate selecting another cover file with appropriate size. As the cover file is a word file with .DOC extension, PowerPoint file with.PPT extension, the notepad file with .TXT extension, or the MATLAB script file with .M extension, therefore the necessity is to first convert the matrix into single dimension, and the secret file is also converted into single dimension. To add to the security of the covert message file, it is converted into binary format and each binary data undergoes right shift, XOR operation, then again right shift and finally XOR operation for those binary bits, Next we need to convert the above binary format secret file into single dimension. Then requirement is to convert the single dimensional cover file data into binary format, Followed by replacement of the 1st bit of the binary cover file data with the one’s complement of the binary format secret file data. Next convert the single dimensional cover file data into decimal format to obtain the stego-image. By taking an illustration this entire development can be with no trouble understood

12

Assume cover file = 1 6 2 7 3 8 4 9 5 10 And Secret file=10 Before Bit exchange the secret file in binary form is 1 0 1 0 0 0 0 0 Now we perform right shift, XOR operation, then again right shift and finally XOR operation on the above secret pattern of binary file and after this Bit exchange the secret file in binary form is 1 0 0 0 0 0 1 0 Take one's complement of the secret message 1 0

0 1

0 1

0 1

0 1

0 1

1 0

0 1

->secret message ->1's complement

Cover file in binary format 1 0 0 0 0 0 0 0 1 0 0 0 0 0 1 1 0 0 0 0 0 0 0 1 0 0 0 0 1 0 1 0 0 0 0 0 1 1 0 0 0 0 1 1 1 0 0 0 0 0 0 0 1 0 0 0 1 0 0 1 0 0 0 0 1 0 1 0 0 0 Now we will embed the secret message at the 1st bit 0 0 0 0 0 0 1 1 0 0 0 0 1 1 0 0 0 0 1 0 1 0 0 0 1 0 1 0 0 0 1 1 1 0 0 0 0 1 1 0 0 0 1 0 0 1 0 0 1 0 0 1 0 0 0 1 0 1 0 0 Thus the stego file =

13

0 :1 0 :2 0 :3 0 :4 0 :5 0 :6 0 :7 0 :8 0 :9 0 :10 0 0 0 0 0 0 0 0 0 0

0 0 0 0 0 0 0 0 0 0

0 7 3 6 3 9 5 9 5 10 At the receiver’s end Decryption process is followed by entering the stego file. The secret file size, password and Pattern are extracted from the cover file. The user needs to enter the password. If the password mismatch, then the decryption is terminated. If the password matches, then the decryption is performed. The decryption process follows the same procedure of the encryption but in the reverse direction. The extracted secret file is stored. Then secret file is displayed.

4 Result The results for the proposed work are investigated with samples of word file with .DOC extension, PowerPoint file with.PPT extension, the notepad file with .TXT extension, and the MATLAB script file with .M extension. In the ensuing four cases (fig 1, 2, 3, 4) of the results acquired in the projected work, the cover media is a word file with the extension .DOC while the secret messages are taken as word file with .DOC extension (fig 1), PowerPoint file with .PPT extension (fig 2), notepad file with .TXT extension (fig 3), and MATLAB script file with .M extension (fig 4) respectively.

Fig 1: Files in use for investigation : cover file of .doc extension and secret file with .doc extension

14

Fig 2: Files in use for investigation : cover file of .doc extension and secret file with .ppt extension

Fig 3: Files in use for investigation : cover file of .doc extension and secret file with .txt extension

Fig 4: Files in use for investigation : cover file of .doc extension and secret file with .m extension

15

Now in next four cases (fig 5, 6, 7, 8) of the results acquired in the projected work, the cover media is a PowerPoint file with the extension .PPT while the secret messages are taken as word file with .DOC extension (fig 5), PowerPoint file with .PPT extension (fig 6), notepad file with .TXT extension (fig 7), and MATLAB script file with .M extension (fig 8) respectively.

Fig 5: Files in use for investigation : cover file of .PPT extension and secret file with .doc extension

Fig 6: Files in use for investigation : cover file of .PPT extension and secret file with .PPT extension

Fig 7: Files in use for investigation : cover file of .PPT extension and secret file with .TXT extension

16

Fig 8: Files in use for investigation : cover file of .PPT extension and secret file with .M extension

The next four cases (fig 9, 10, 11, 12) of the results acquired in the projected work, the cover media is a notepad file with the extension .TXT while the secret messages are taken as word file with .DOC extension (fig 9), PowerPoint file with .PPT extension (fig 10), notepad file with .TXT extension (fig 11), and MATLAB script file with .M extension (fig 12) respectively.

Fig 9: Files in use for investigation : cover file of .TXT extension and secret file with .DOC extension

Fig 10: Files in use for investigation : cover file of .TXT extension and secret file with .PPT extension

17

Fig 11: Files in use for investigation : cover file of .TXT extension and secret file with .TXT extension

Fig 12: Files in use for investigation : cover file of .TXT extension and secret file with .M extension

Now in next four cases (fig 13, 14, 15, 16) of the results acquired in the projected work, the cover media is a MATLAB script file with the extension .M while the secret messages are taken as word file with .DOC extension (fig 13), PowerPoint file with .PPT extension (fig 14), notepad file with .TXT extension (fig 15), and MATLAB script file with .M extension (fig 16) respectively.

Fig 13: Files in use for investigation : cover file of .M extension and secret file with .DOC extension

18

Fig 14: Files in use for investigation : cover file of .M extension and secret file with .PPT extension

Fig 15: Files in use for investigation : cover file of .M extension and secret file with .TXT extension

Fig 16: Files in use for investigation : cover file of .M extension and secret file with .M extension

The main motive of the reserech work to embed information in such a way that embedded data leave no traces or signs of Steganography use is been very efficiently

19

achieved as seen in the resultant set if images. The success of the proposed technique may be viewed by the complexity and difficulties involved in detecting or distinguishing the hidden data in the carrier.

5 Conclusion In the presented work, the chief essential matter is the discretion and secrecy of the buried, confidential message data being communicated. The proposed technique presents the improvement to enclose whichever language; it is a advancement in the enrichment of the Steganography practices.This algorithm harmony is with the text Steganography and is pretty tremendously dexterous as here it is tough and tricky to figure out the authenticate information that is veiled at the back. The outcomes for the intended work are attained using the MATLAB version 7.10.0.499(R2010a).

References 1.

Mohammad Shirali-Shahreza, “Text Steganography By Changing Words Spelling”, Feb. 17-20, 2008 Icact 2008, Isbn 978-89-5519-136-3.

2.

M. Hassan Shirali-Shahreza and Mohammad Shirali-Shahreza, “Arabic/Persian Text Steganography Utilizing similar Letters With Different Codes”, The Arabian Journal For Science And Engineering, Volume 35, Number 1b, April 2010. L. Y. Por, B. Delina, Information Hiding: A New Approach In Text Steganography, 7th Wseas Int. Conf. On Applied Computer & Applied Computational Science (Acacos '08), Hangzhou, China, April 6-8, 2008, Isbn: 978-960-6766-49-7, Issn: 1790-5117. Nuur Alifah Roslan, Ramlan Mahmod, Nur Izura Udzir, “Sharp-Edges Method In Arabic Text Steganography”, Journal Of Theoretical And Applied Information Technology 15th November 2011. Vol. 33 No.1, Issn: 1992-8645, E-Issn: 1817-3195. Mayank Srivastava, Mohd. Qasim Rafiq, and Rajesh Kumar Tiwari, “A Novel Approach to Hindi Text Steganography”, klkdV.V. Das, J. Stephen, and Y. Chaba (Eds.): CNC 2011, CCIS 142, pp. 295–298, 2011.© Springer-Verlag Berlin Heidelberg 2011. Indradip Banerjee, Souvik Bhattacharyya, Prof. Gautam Sanyal, “Novel Text Steganography through Special Code Generation”, International Conference on Systemics, Cybernetics and Informatics, 2011 J.T. Brassil, S. Low, N.F. Maxemchuk, and L.O’Gorman, "Electronic Marking and Identification Techniques to Discourage Document Copying", IEEE Journal on Selected Areas in Communications, vol. 13, Issue. 8, October 1995, pp. 1495-1504. J. T. Brassil, S. Low, and N. F. Maxemchuk, “Copyright protection for the electronic distribution of text documents,” Proceedings of the IEEE,vol. 87, no. 7, pp. 1181–1196, July 1999. A. Gutub and M. Fattani, "A Novel Arabic Text Steganography Method Using Letter Points and Extensions", Proceedings of the WASET International Conference on Computer,Information and Systems Science and Engineering (ICCISSE), Vienna, Austria, Vol.21, 2007, pp. 28-31.

3.

4.

5.

6.

7.

8.

9.

20

Probability Inclusion Degrees and Probability Fuzzy Attribute Implications for Interval-Valued Fuzzy Sets Yanhui Zhai1 , Deyu Li1 , and Kaishe Qu1 School of Computer and Information Technology, Shanxi University, Taiyuan 030006, China

Abstract. In this paper, we point out that the definition of fuzzy attribute implication actually implies a minimal degree and thus may be impractical in some applications. To solve this, we introduce probability information to fuzzy inclusion degree, and discuss some properties of this definition. Illustrations verifies the correctness of probability fuzzy attribute implication. Keywords: probability fuzzy attribute implication; probability inclusion degree; fuzzy attribute implication; interval-valued fuzzy formal context; residuated lattice

1

Introduction

As an extension of Zadeh-style fuzzy set [1], interval-valued fuzzy set introduced by [2] exhibits a subinterval of [0, 1], instead of a fixed value, for an object as its membership degree. This approach indeed introduces uncertainty into fuzziness and shows more flexibility than classical fuzzy set. Subsequently, interval-valued fuzzy set have been studied from various viewpoints [3–5] and also on the relationships with other extensions of fuzzy set [6]. In Formal Concept Analysis (FCA) [7], interval-valued fuzzy set was first introduced by Burusco in [8] as interval-valued formal contexts (IVFF contexts). On the other hand, in FCA, the most useful framework developed by Belohlavek [9] is based on complete residuated lattice [10, 11]. In this framework, Belohlavek presented the semantical and syntactical characteristics of fuzzy attribute implications. In this paper, in order to apply these results into IVFF contexts, in the light of [12] we will show that the study of interval-valued fuzzy set can be fitted into the framework. Moreover, we state that the definition of fuzzy attribute implication (FAI) in fact implies a minimal degree and thus may be impractical in some applications. To produce a more reasonable result, this paper introduces probability information to fuzzy attribute implication, and then discusses some properties of this definition.

21

2

Yanhui Zhai, Deyu Li, and Kaishe Qu

2

Interval-valued Fuzzy Sets: A Perspective of Residuated Lattice

An interval-valued fuzzy set A in a universe U is given by A(u) = [µ(u), µ(u)], where µ, µ : U → [0, 1] are the so-called lower and upper membership functions such that for any u ∈ U , 0 ≤ µ(u) ≤ µ(u) ≤ 1. That is, the degree to which an element necessarily belongs to A equals to µ, while the degree to which the element possibly belongs to A equals to µ. The constrained condition of µ(u) ≤ µ(u) means that necessity implies possibility. It is easy to see that Zadahstyle fuzzy set is a special case of interval-valued fuzzy set when µ = µ. In this paper, we also call the values of µ(u)and µ(u) lower and upper approximations of µ respectively. Interval-valued fuzzy set can also be constructed based on a complete lattice L, just by defining µ, µ : U → L. Moreover, if L is a complete residuated lattice [10, 11] we can take interval-valued fuzzy set as a special L-fuzzy set as shown below. To this end, we recall the definition of complete residuated lattice. The following lemma can be used to connect interval-value fuzzy sets with residuated lattice[10]. Lemma 1. [B. V. Gasse, etc.[12]] For a residuated lattice L, define the sublattice T (L) of L × L as T (L) = {(x1 , x2 ) ∈ L × L|x1 ≤ x2 }, and then we can generate a residuated lattice L = (T (L), ∧, ∨, ⊗, →, 0, 1) such that – – – – –

(a1 , b1 ) ∧ (a2 , b2 ) = (a1 ∧ a2 , b1 ∧ b2 ), (a1 , b1 ) ∨ (a2 , b2 ) = (a1 ∨ a2 , b1 ∨ b2 ), (a1 , b1 ) ⊗ (a2 , b2 ) = (a1 ⊗ a2 , b1 ⊗ b2 ), (a1 , b1 ) → (a2 , b2 ) = ((a1 → a2 ) ∧ (b1 → b2 ), b1 → b2 ), 0 = (0, 0) and 1 = (1, 1)

for (a1 , b1 ), (a2 , b2 ) ∈ T (L). Using the above lemma, one can observe that each interval-valued fuzzy set A can then be viewed as a mapping A : U → T (L), and thus is a special case of L-fuzzy set.

3

Probability Inclusion Degree

For data table, entries can also be interval values, called interval-valued fuzzy formal contexts (IVFF contexts for short) [8, 13]. Formally, an IVFF contexts is a triple hX, Y, Ii, where X is a set of objects, Y is a set of attributes, and I is a mapping X × Y → T (L). An example concerning interval values is taken from [14, 6] and shown in Figure 1. So far we have considered how to apply residuated lattice structure to intervalvalued fuzzy set, and thus the study of IVFF contexts can be subsumed into the framework of Belohlavek [15]. However, we note here that in some cases this framework may be impractical as shown below.

22

Probability Inclusion Degrees and Probability Fuzzy Attribute Implications

3

Table 1. Fats and oil data after transformed Objects Linseed oil Perilla oil Cottonseed oil Sesame oil Camellia oil Olive oil Beef tallow Hog fat

Specific gravity [0.930,0.935] [0.930,0.937] [0.916,0.918] [0.920,0.926] [0.916,0.917] [0.914,0.919] [0.860,0.870] [0.858,0.864]

Freezing point [0.000,0.138] [0.338,0.354] [0.323,0.400] [0.323,0.354] [0.092,0.185] [0.415,0.508] [0.877,1.000] [0.754,0.908]

Iodine value Saponification value [0.774,0.976] [0.000,0.929] [0.905,1.000] [0.833,0.940] [0.351,0.435] [0.845,0.952] [0.381,0.452] [0.821,0.893] [0.238,0.250] [0.845,0.893] [0.232,0.298] [0.821,0.929] [0.000,0.048] [0.857,0.964] [0.077,0.220] [0.857,1.000]

To start with, in this paper we restrict our discussion to the special residuated lattice [0, 1] and denote by LU the collection of all [0, 1]-fuzzy sets. Then, recall that the study of FAIs developed by Belohlavek in [15] depended basically upon the inclusion degree (i.e., subsethood degree) between fuzzy sets (or L-sets): S(A, B) =

∧

(A(u) → B(u)),

u∈U

for two fuzzy sets A and B on the universe U . In other words, S(A, B) equals to the minimal degree among all the degrees to which elements u belong to B whenever they belong to A. In some applications, this lower approximation may be correct and practical, when we want to rank fuzzy sets strictly. However, in others (for example, data tables may contain noise and the minimal degree is obtained accidentally on some noise object), this would be impractical. For example, let A, B be two fuzzy sets on interval [0, 1] with A(u) ≤ B(u) for any u ∈ U except some w ∈ U for which A(u) = 1 and B(u) = 0. In this case, S(A, B) = 0. This result would be unacceptable, since A is almost everywhere contained in B. Therefore, to make it more practical, it is necessary to add the measure information to the definition of S. Then, an intuitive method is to add probability information to the definition: SP (A, B) =

∑

((A(u) → B(u))P (u)).

u∈U

Obviously, SP is more acceptable than S in this case. ∨ Defining SM (A, B) = u∈U (A(u) → B(u)), we have the following boundaries for probability inclusion degree. Theorem 1. For any A, B ∈ LU , we have SP (A, B) ≤ SM (A, B). Proof. It is straightforward.

23

∧

u∈U (A(u)

→ B(u)) = S(A, B) ≤

4

4

Yanhui Zhai, Deyu Li, and Kaishe Qu

Probability Fuzzy Attribute Implications

In this section, we consider how to introduce probability information to FAIs. First, recall the definition of FAI [9]. An FAI over a finite set Y of attributes is an expression A ⇒ B, where A, B ∈ LY . Then Belohlavek [9] defined a degree kA ⇒ BkM ∈ L for a fuzzy set M ∈ LY by kA ⇒ BkM = S(A, M )∗ → S(B, M ). Using probability inclusion degree instead, we define the probability FAI degree as ∗ kA ⇒ BkP M = SP (A, M ) → SP (B, M ). For a given interval-valued data table hX, Y, Ii, the probability valid degree to which A ⇒ B is valid in hX, Y, Ii is given by P kA ⇒ BkP hX,Y,Ii = kA ⇒ Bk{Ix |x∈X} ,

where Ix represents an object fuzzy set (e.g. the “Perilla oil fuzzy set”), while Belohlavek defined it by kA ⇒ BkhX,Y,Ii = kA ⇒ Bk{Ix |x∈X} .

5

Rationality of Probability FAI

To illustrate the preference of our definitions, we obtain the most distinct pair of kA ⇒ BkhX,Y,Ii = [0.1466, 0.1466] and kA ⇒ BkP hX,Y,Ii = [0.7866, 0.7866] according to lower and upper approximations, and list the premise and consequence of the FAI in Table 2. Table 2. Premise and Consequence of FAI A FAI No. Specific gravity Freezing point Iodine value Saponification value 1 [0.1540,0.3134] [0.6745,0.9910] [0.0428,0.1305] [0.5269,0.7404] 2 ... ... ... ... B FAI No. Specific gravity Freezing point Iodine value Saponification value 1 [0.3365,0.7288] [0.6526,0.9344] [0.7605,0.9839] [0.3159,0.5028] 2 ... ... ... ...

To obtain this result, we show the FAI degrees and probability FAI degrees in Table 3.

24

Probability Inclusion Degrees and Probability Fuzzy Attribute Implications

5

Table 3. FAI degrees and probability FAI degrees Objects Linseed oil Perilla oil Cottonseed oil Sesame oil Camellia oil Olive oil Beef tallow Hog fat

kA ⇒ BkIx [1.0000,1.0000] [1.0000,1.0000] [0.9420,1.0000] [0.9720,1.0000] [1.0000,1.0000] [0.7301,0.7970] [0.1466,0.1466] [0.3165,0.3190]

kA ⇒ BkP Ix [1.0000,1.0000] [1.0000,1.0000] [0.8769,0.8769] [0.8812,0.8812] [0.8307,0.8307] [0.8427,0.8427] [0.7866,0.7866] [0.8232,0.8232]

According to Table 3, both of values of kA ⇒ BkhX,Y,Ii and kA ⇒ BkP hX,Y,Ii are determined by object “Beef tallow”, i.e., the fuzzy set M = Beef tallow ={[0.860, 0.870]/Specific gravity, [0.877, 1.000]/Freezing point, [0.000, 0.048]/Iodine value, [0.857, 0.964]/Saponification value}. In fact, we have kA ⇒ BkM = [0.2823 ∧ 0.1466, 0.1466] kA ⇒ BkP M = [0.8206 ∧ 0.7866, 0.7866], Then, the second components decide the values of kA ⇒ BkM and kA ⇒ BkP M, and the most distinct parts are S(B, M ) = [0.0641, 0.0641] and SP (B, M ) = [0.7660, 0.7660]. Thus the analysis of FAI valid degrees is then reduced to the differences of S(B, M ) and SP (B, M ). As shown before, the former is assigned to the minimal degree and the later is a weighted average. Accordingly, if one do not want the FAI valid degree determined by some objects that perhaps contain noise, probability FAI valid degree will be proper and required.

6

Conclusion and Further Work

In this paper, we considered probability inclusion degree and probability FAI. We first show that the study of interval-valued fuzzy set can be fitted into the framework developed by Belohlavek. Second, we state that the definition of FAI by Belohlavek in fact implies a minimal degree and thus may be impractical in some applications. Thus this paper introduces probability information to fuzzy inclusion degree and then to fuzzy attribute implication, and discuss some properties of this definition. The result shows their differences and verifies the correctness of probability fuzzy attribute implication in some illustrations. One of shortcomings of the paper is that we limit our discussion only to the interval [0, 1] and fail to give further results on general residuated lattice, due to the fact that one can not directly give meaningful operators between probability measure and residuated lattice. One approach to solve this problem is to consider

25

6

Yanhui Zhai, Deyu Li, and Kaishe Qu

a probability ∨ measure P : U → L, which is defined on residuated lattice L and such that u∈U∧ P (u) = 1. And then define the probability inclusion degree as SP (A, B) = u∈U (P (u) → (A(u) → B(u))), which however is still under investigation.

Acknowledgments The work is supported by National Natural Science Foundation of China (61070100, 60970014) and the Natural Science Foundation of Shanxi, China (2010011021-1).

References 1. Zadeh, L.A.: Fuzzy sets. Information and Control 8(3) (1965) 338–353 2. Sambuc, R.: Fonctions Φ-floues. PhD thesis, Univ. Marseille (1975) 3. Kreinovich, V., Zuo, Q.: Interval-valued degrees of belief: Applications of interval computations to expert systems and intelligent control. International Journal of Uncertainty, Fuzziness and Knowledge-based Systems 5 (1997) 317–358 4. Moore, R., Lodwick, W.A.: Interval analysis and fuzzy set theory. Fuzzy Sets and Systems 135 (2003) 5–9 5. Dubois, D., Prade, H.: Interval-valued fuzzy sets, possibility theory and imprecise probability. In: European Society for Fuzzy Logic and Technology. (2005) 314–319 6. Denoeux, T., hlne Masson, M.: Multidimensional scaling of interval-valued dissimilarity data. Pattern Recognition Letters 21 (2000) 83–92 7. Ganter, B., Wille, R.: Formal concept analysis: mathematical foundations. Springer (1999) 8. Burusco, A., Fuentes-gonzlez, R.: The study of the interval-valued contexts. Fuzzy Sets and Systems 121 (2001) 439–452 9. Belohlavek, R., Vychodil, V.: Attribute implications in a fuzzy setting. In: ICFCA. (2006) 45–60 10. H´ ajek, P.: Mathematical of Fuzzy Logic. Kluwer (1998) 11. Belohlavek, R.: Fuzzy Relational Systems: Foundations and Principles. Kluwer Academic Publishers, Norwell, MA, USA (2002) 12. Gasse, B.V., Cornelis, C., Deschrijver, G., Kerre, E.E.: Triangle algebras: Towards an axiomatization of interval-valued residuated lattices. In: Rough Sets and Current Trends in Computing. (2006) 117–126 13. Djouadi, Y., Prade, H.: Interval-valued fuzzy formal concept analysis. In: International Syposium on Methodologies for Intelligent Systems. (2009) 592–601 14. Ichino, M., Yaguchi, H.: Generalized minkowski metrics for mixed feature-type data analysis. IEEE Transactions on Systems, Man, and Cybernetics 24 (1994) 698–708 15. Belohl´ avek, R., Vychodil, V.: Fuzzy attribute logic: attribute implications, their validity, entailment, and non-redundant basis. In: Proc. IFSA 2005 World Congress. (2005) 622–627

26

Analytical Evaluation of MDPSK and MPSK Modulation Techniques over Nakagami Fading Channels Alam S. M. Shamsul1, Kwon GooRak2, and Choi GoangSeog3 Department of Information and Communication Engineering, Chosun University 309 Palmun St. Dong, GwanJu,501-759, Korea Choi GoangSeog is a corresponging author Email: 1{[email protected]}, 2{[email protected]}, 3{[email protected]}

Abstract. To design the wireless communication channel, it is required to know the statistical model of channel accurately. Nakagami distribution is extremely convenient tool for analyzing the performance of wireless fading channels. In this paper, a unified analytical framework is presented to obtain the closed-form solutions for calculating the average symbol error rates of MDPSK and coherent MPSK modulation, with or without diversity reception over Nakagami–m fading channels. Result shows that an error rate of the fading channel typically depends on Nakagami parameters (m) and the space diversity (N). The fading effect of channel is inversely proportional to m and N that represents, for high value of m, the channel acts like AWGN transmission and the low value of N creates high fading environments. Also this paper portrays a comparison between MPSK and MDPSK modulation formats. The result proves that MPSK demonstrates better performance compared to MDPSK over any fading and non-fading conditions. Keywords: Fading channels, M-ary Differential Phase-Shift Keying (MDPSK), M-ary Phase-Shift Keying (MPSK), Nakagami-m, Probability Density Function (PDF), Symbol Error Rate (SER)

1 Introduction In mobile communication systems, the propagation of signal is characterized by three Non Line Of Sight (NLOS) conditions. These are reflection, diffraction, and scattering which roughly illustrate the radio propagation by three other phenomena like path loss, shadowing and multipath fading. Shadowing and multipath fading are known as long term and short term fading respectively. These two phenomena can be statistically described by the fading models [1]. The analytical methods of this paper are applied for calculating the SER of both a coherent M-ary phase shift keying (MPSK) system and an M-ary differential phase shift keying (MDPSK) over slow and flat Nakagami-m fading channel. Nakagami distribution sometimes is considered as good in term of flexibility and accuracy than Rayleign, Rician or log-normal distributions [2]. Several statistical distributions have been proposed by the researchers for channel modeling of fading envelope under short and long term fading conditions. Short-term fading models include the Rayleigh, Weibull, Rice, and Nakagami distributions [1].

27

Turin [3] proposed one model that associates each path with a set of three random variables like the strength, the modulation delay and the carrier phase shift [2]. When the delay is unknown, the fading is followed as Nakagami distribution, which seems to perform better than Rice distribution [2]. Also Fotohabadi et. al. [7] showed that the consumed power in Nakagami sub-channel is less than that in Rayleigh subchannel. George et. al. [1] investigated the suitability of Nakagami distribution to approximate the lognormal distribution. Juang et. al. [8] presented a method for parameter selecting of a finite-state Markov chain to match with Nakagami-m fading channel. In this paper, the performances of different modulation techniques are simulated over Nakagami-m fading channel.

2 Statistical Characteristics of Nakagami-m fading channel The PDF of a received signal under Nakagami fading with parameters given by:

where

and Ω is

is the Gamma function, a is strength of received signal,

and

. The Nakagami distribution becomes Rayleigh if the value of m=1. A Nakagami distribution (m, ) can approximate a Rice distribution [2] with m 1. For a given distribution, we shall define the amount of fading (AF) is AF = var {a2} / [E{a2}] 2. For the Nakagami distribution, and [2]. In particular , then AF = 0 which corresponds to a situation of “no fading”; case when for , AF = 1 which corresponds to “Rayleigh fading” and for , AF = 2, which is the severest fading assumed by Nakagami fading.

3 SER derivation of MPSK over Nakagami-m fading channels It can be assumed that there are N diversity channels carrying the same information bearing signal. Each channel is modeled as frequency non-selective slowly Nakagami-m fading channel corrupted by additive white Gaussian noise (AWGN) process. In fading, instantaneous SNR becomes a random variable and assume that the channel is identical, the PDF of at the output of the maximum ratio combiner (MRC) may be written as [4], (1)

28

where, G (.) is the gamma function, m is the Nakagami fading parameter, which is assumed to be identical for all channels and is the average SNR associated with as . If the PDF of ( ) is known, then the each symbol, which is related to average SER in fading can be calculated by the following equation (2). (2) In here, is the conditional probability of symbol error over AWGN channel. The SER of M-PSK is given in the exponential form and can be written as [5], (3) Equation (4) represents the integral relationship explained in [6]. (4) By substituting equations (1) and (3) in equation (2), and using the relationship of equation (4), the average SER over Nakagami fading channel with N order diversity is,

(5) Equation (5) represents the average SER of coherent M-PSK over Nakagami-m fading channels. If m = ∞, equation (5) represents the equation (3) which is the SER of MPSK over AWGN channel.

4 SER Derivation of MDPSK over Nakagami-m Fading Channels It is shown in [5] that the probability of symbol error of M-DPSK over AWGN channel is given as (6)

Substituting the values from (6) and (1) into equation (2), and using the relation of equation (4), the average symbol error probability of MDPSK over Nakagami fading channel is,

(7)

29

5 Evaluation of MPSK over Nakagami-m Fading Channel Equation (5) shows the exact numerical represent of SER of M-PSK modulation technique over Nakagami-m fading channel. In here, represents the average signal to noise ratio in dB and N shows the number of available channels. So SER over Nakagami-m fading channel can be plotted by varying the channel numbers (N) and the number of symbols M for PSK modulation. If the value of m = 1, it shows Rayleigh fading channel and it represents AWGN channel for m >> 1. Interestingly it can be shown that the increasing value of m coverts the channel from Nakagami to AWGN. Figure 1 depicts the average symbol error rate of BPSK (M = 2) modulation. In this figure, number of channel is one (N = 1) and m is varied from 0.5 to 106 (theoretically infinity). From this figure, it is easily showed that when m = 0.5, its performance is worse and when m is gradually increased then the channel converts to non fading channel like AWGN. But, here the overall performance of fading channel is not so good because it is plotted for N = 1 (no diversity). Figure 2 shows the symbol error rate for BPSK modulation with multiple channel numbers (N = 2, 12). It is plotted for same conditions as figure 1 except diversity. Here diversity is applied (N> 1) and it shows good performance compared to figure 1. If the fading effects are gradually reduced then channel goes close to AWGN channel. For example from these two figures, it can be shown that for Nakagami channel (m = 2) SER is almost near to 10-3at 15 dB SNR for without diversity condition but when diversity is applied then SER is reduced and almost equals to 10-5.

Fig. 1. SER of BPSK for N = 1 (without diversity) over different fading channels

Fig. 2. SERof BPSK for N = 2, 12 (with diversity) over different channels

Figure 3 is simulated for different symbol values of PSK (M = 4, 16, 64) modulation technique including diversity and without diversity. For severe fading channel (m = 0.5), figure 3 shows that, when number of symbol is low (M = 4), the performance is high. But as the fading parameter m improves from 0.5 to 1, there is significant improvement of response curve shown in figure 3 for without diversity conditions. Similarly, from figure 3, low symbol value of PSK shows good performance because if the phase variation of carrier signal is low then the phases of carrier are less corrupted by channel. Comparing the different curves in figure 3, it can be said that the vulnerability increases with M. For example, at a given SNR, the SER is raised least for M = 4 and greatest for M = 64.

30

Fig. 3. SER of M (= 4, 16, 64) PSK for N = 1 Fig. 4. SER of DBPSK over AWGN, (without), N = 12(with diversity) over Rayleigh, and Nakagami-m channel Nakagami (m = 0.5) channel

Fig. 5. SER of M = 4, 16, 64 DPSK for N = 1 (without diversity), N = 12 (with diversity) over Nakagami fading Channel (m = 5)

Fig. 6. SER of M = 4, 16, 64 for DPSK and MPSK over Nakagami fading Channel

6 Evaluation of MDPSK over Nakagami-m Fading Channel Equation (7) shows the exact numerical presentation of symbol error rate for MDPSK over Nakagami-m fading channel. Figure 4 is simulated for DBPSK over different fading channels without diversity condition but diversity is applied for only Nakagami parameter m = 0.5. From this figure, it can be said that when diversity is applied then the fading channels show good result compared to without diversity condition. Figure 5 shows the performance for M = 4, 16, 64 DPSK in various symbols (different M) including with and without diversity. For fixed value of fading parameter m = 5, M = 4 DPSK shows good result compared to other values of M. Now if we compare the result between MPSK and MDPSK over Nakagami fading channel then MPSK is an attractive communication technique and shows very good performance than MDPSK modulation technique. Figure 6 shows such simulation result for M = 4, 16, 64 DPSK and PSK over Nakagami fading channel. From the above simulation curves, it can be written that there should be a tradeoff between fading parameter m and diversity N. If channel behaves like strong fading effects (for low value of m) then it is wise to increase diversity. Similarly if there is an AWGN channel (m = ∞) then it is not

31

necessary to apply diversity. Therefore, in nutshell, number of symbols (M) and SNR are significant parameters to design efficient communication system. The increasing tendency of M consumes large SNR at the given error rate. It is not always possible to transmit data with high rate. Depending on the channel response, designers will assign the symbol rate by observing the error responses.

7 Conclusions We outline analytical solutions for computing average SER of MDPSK and MPSK over Nakagami fading channel. These are simulated and compared for various values of fading parameter m. In this paper, PSK modulation is widely analyzed for different fading channels using space diversity. For that reason engineers can get useful information to evaluate and design the radio communication link over Nakagami fading channel.

References 1. Karagiannidis, G.K., Sagias, N.C., Mathiopoulos, T.: The N *Nakagami Fading Channel Model. In: International Symposium on Wireless Communication Systems, pp.185 - 189. Italy (2005) 2. Charash, U.: Reception Through Nakagami Fading Multipath Channels with Random Delays. IEEE transactions on Communication, vol. 27(4), pp. 657-670. (1979) 3. Annamalai, A., Tellambura, C., Bhargava, V.: A General Method for Calculating Error Probabilities over Fading Channels. IEEE transactions on Communication, vol. 53 (5), pp. 841 – 852. (2005) 4. Patterh, M.S., Kamal, T.S., Sohi, B.S.: Performance of coherent square M-QAM with Lth order diversity in Nakagami-m fading environment. IEEE Conference on Vehicular Technology, vol. 6, pp.2849-2853. (2000) 5. Alouini, M.S., Goldsmith, A.J.: A Unified Approach for Calculating Error Rates of Linearly Modulated Signals over Generalized Fading Channels. IEEE Transaction on Communication, vol. 47(9), pp. 1324 – 1334. (1999) 6. Gradshteyn, I.S., Ryzhik, I.M.: Table of Integrals, Series, and Products. Academic Press, pp. 346.USA (2007) 7. Fothohabad, V., Said, F.: Comparison of the Rayleigh and Nakagami fading channels MIMO multicarrier system. In Conference on Wireless Advanced, pp. 295-300. London (2011) 8. Juang, M.A., Parsley, M.B.: New results on finite-state Markov models for Nakagami fading channels. In Conference on Military Communication, pp. 453-458. Baltimore (2011)

32

A Fuzzy Clustering Approach for Missing Value Imputation with Non-Parameter Outlier Test Jing Tian1, Bing Yu2, Dan Yu3 and Shilong Ma4 State Key Laboratory of Software Development Environment Beihang University 1 Corresponding Author: [email protected] { 2yubing;3yudan;4slma }@nlsde.buaa.edu.cn

Abstract. Missing value is a challenging issue in data mining, as information deficiency negatively affects both data quality and reliability. This paper focuses on an algorithm of a fuzzy clustering approach for missing value imputation with noisy data immunity. The PCFKMI (Pre-Clustering based Fuzzy K-Means Imputation) method aggregates data instances to more accurate clusters for further appropriate estimation via information entropy after resampling pre-clustering and outlier test. Experimental results demonstrate that the PCFKMI proposed obtains higher precision both on quantitative and on nominal attributive missing value completion than other classic methods under all missingness mechanisms at varying missing rates with abnormal values. Keywords: Fuzzy clustering, Missing value imputation, k-Means, Outlier Test

1

Introduction

Data sets usually contain errors in the form of missing values, as massive various collection systems are unable to fully guarantee the successful information access as avoiding omission and no response. Since those data with missing parametric values undermine the determinacy, it is difficult to extract useful knowledge from them [1]. Therefore, many strategies have been designed to manage the missing data to promote both data quality and reliability. Currently, these methods are categorized as follows [2, 3] : a) Ignoring. It is the simplest way to directly delete the data records or items with missing values, including List-wise Deletion and Pair-wise Deletion. However, the conduction drops abundant values. b) Toleration. The interior tactic requires direct data mining on the set with fragmentary data. But as most analytic procedures work on integrated dataset, it is not widely applicable. c) Imputation. The kernel fashion is to seek plausible values using known information. Whereas the most probable values could be estimated to renew a complete data set considering reasonable models, it is proved to be the best means to process missing values. Generic k-Means becomes prevalent because of the splendid efficiency and accuracy, but the defects are still obvious [4]: a) the imperfect data may disrupt the final result due to the entire instances involvement; b) some parameters, such as the total number of clusters, need beforehand settings; c) the similarity metric among

33

1

instances is simply measured by Euclidean distance; d) the initial randomly-selected seeds result in uncertainty; e) outliers impair the precision. The main contributions the paper has made are: a) source data is separated into complete set and incomplete set to dispel the disturbance of missingness on clustering; b) the well-partitioned clusters and initial centroids are gained by evaluating the hierarchical clustering on pre-collected samples; c) fuzzy k-Means is used during clustering; d) outliers are properly tackled once detected; e) similarity metric and membership functions are designed for better classification and estimation. The rest of the paper is organized as follows: Section 2 reviews the related work. Section 3 states the basic notations. Section 4 elaborates the procedures of the PCFKMI algorithm. Section 5 conducts empirical experiments on benchmark dataset. Finally, Section 6 summarizes the paper.

2

Related Work

Little and Robin [5] defined three types of missingness mechanisms: a) Missing Completely At Random (MCAR). For each individual, the missingness is independent both of observable and unobservable values, mathematically written as Pr(r|Xc,Xic)=Pr(r). b) Missing At Random (MAR). It is equivalent to saying that given the observed data, the missingness does not depend on the unobserved data, namely Pr(r|Xc,Xic)=Pr(Xc). c) Not Missing At Random (NMAR). When neither MCAR nor MAR holds, the data are NMAR. Then the principal approaches are roughly classified as follows: a) Mean/Mode Substitution: It replaces the missing values by the mean or mode of all the observations or a subgroup at the same variable. b) Hot-deck imputation: It completes missing data with comparable data from the same set. It is a means of imputing data, using the data from other observations in the sample at hand [6]. c) Regression imputation: Regression models such as linear regression and logistic regression etc. are used to predict missing values [7]. d) EM imputation: It is based on ExpectationMaximization (EM) algorithm [8]. The iterative procedure is employed to calculate the sufficient statistics and estimate the parameters. e) Multiple imputation: It substitutes each missing value with two or more plausible values that represent the uncertainty about the right value to impute [9]. Each of the two or more resulting complete data sets is then analyzed using standard complete-data methods. All the analysis becomes combined to reflect both the inter- and intra-imputation variability. f) Machine learning based imputation: It acquires the features of interested unknown data by behavior evolution after sample data processed. The essence is to automatically learn samples for complicated pattern cognition and to intelligently predict the missing values [10]. To examine the performance, researchers have done substantial work. Magnani [11] overviews the main strategic methods about dealing with missing data, including traditional methods, global imputation, local imputation and parametric estimation etc. and comments on their advantages and disadvantages. In his view, statistical methods have been proved very efficient but they demand strong model hypothesis. Batista and Monard put forward 10-NNI, and conclude that the method could predict both

34

2

quantitative and qualitative data in comparison of mean substitution [12], C4.5 and CN2. Since it is in lazy mode, no explicit models are needed. However, the defect is that it merely works on large dataset under MCAR. Zhang et al. [13] present a clustering-based imputation method, which the prediction of incomplete instance A is generated by the complete instance that the most similar with. Nevertheless people pay less attention to the impact that noisy data does on the missing value imputation. Van Hulse and Khoshgoftaar analyze the performance of five popular imputation techniques on noisy software measurement data, namely mean imputation, regression imputation, REPTree imputation, Bayesian multiple imputation, and kNN imputation [14] . It concludes that noise has a dramatic negative impact on the effectiveness of these techniques. Furthermore, Bing Zhu et al. [15] examine the impact of noise on imputation methods in a more general situation with experiments on 9 benchmark datasets from different domains. The results signify that the proposed RIBG outperforms other competing methods in the presence of noise.

3

Preliminaries

X={x1,x2,…,xm} is a dataset including m instances, where each owns n attributes, and xij (1 ≤ i ≤ m ,1 ≤ j ≤ n ) , denotes the j-th value of the i-th instance. Then: a) If { x ij ≠ ∅ | ∀1 ≤ j ≤ n} , then xi is a complete instance, denoted as x ic . b) If { x ij = ∅ | ∃1 ≤ j ≤ n} , then xi is an incomplete instance, denoted as x im is . c) If X ck ⊆ X ,1 ≤ k ≤ m and X ck = { x ic },1 ≤ i ≤ m , then X ck is a complete instances set of X. d) If X ick ⊆ X ,1 ≤ k ≤ m and X ick = { ximis },1 ≤ i ≤ m , then X ick is an incomplete instances set of X. e) X c = ∪ X ck is the complete set of X. f) X ic = ∪ X ick is the incomplete set of X.

4

The PCFKMI Algorithm

Initially, the source data set X is split into two sets- Xc and Xic, where Xc ∩ Xic= ∅ and Xc ∪ Xic=X. 4.1

Clustering

P (p ≤ m) instances of Xc are randomly-selected to build a fuzzy similarity matrix R c

[Xcs=( xi )p×n]. Here, seven data records are grasped to exemplify the common scenario of dynamic direct clustering based on the fuzzy similarity matrix [16] in Fig. 1.

35

3

1   0 .9,1     0 .2, 0 .9,1    (a) R =  0 .3, 0 .3, 0 .5,1   0 .4, 0 .4, 0 .1, 0 .6,1     0 .2, 0 .2, 0 .3, 0 .4, 0 .2,1   0 .2, 0 .3, 0 .2, 0 .2, 0 .3, 0 .7 ,1   

(b)

Fig. 1. An example of a fuzzy similarity matrix (a) and the complete weighted graph (b)

The distinguishing levels- λ s and the maximal tree of G are then achieved in Fig. 2.

(a)

(b)

Fig. 2. An example of the maximal tree (a) and the dendrogram of agglomerative clustering (b)

36

4

The result of pre-clustering on sub-sampling is appraised by Fuzzy Pseudo FStatistic (FPFS) [16] owing to the ambiguous margin of inter- and intra-cluster based on fuzzy clustering. Furthermore, for the m*n sample, n n FPFS = [tr (Sinter )(m − c)] [tr (Sintra )(c −1)] is influenced by three matrixes, namely, fuzzy inter-, intra-cluster and mixed scattering matrixes. They are respectively presented c

m

c

n = ∑ ∑ µ ikf ViVi T S inter

as

i =1 k =1

c

m

n = ∑∑ µikf ( X k − Vi )( X k − Vi )T Sintra

,

and

i =1 k =1

m

n Smix = ∑∑ µikf X k X kT

,

i =1 k =1

m

where V = ∑ ( µ ) f x i ik k k =1

m

∑µ

f ik

k =1

, i = 1, 2,..., c , µ ik = 1

2

c

∑ [d (v , x i

k

) d ( v j , x k )] f −1 and

j =1

d( · , · ) is the distance between two instances. Logically, larger FPFS-value explicates better inter-cluster separation and intra-cluster cohesion. So the largest FPFS-value gets the best number of clusters. The centroids are initialized by the result of pre-clustering by v j = ∑ x i | c j | (1). Subsequently, the membership function is introduced to xi ∈ c j

depict the degree of xi subordinate to the k-th cluster by eq. (2), where K

apparently

∑ U (v

j

, xi ) = 1 .

j =1

K

∑ d (v

U ( v k , x i ) = d ( v k , x i ) − 2 / ( f −1)

j

, x i ) − 2 / ( f −1) ; f > 1, i = 1, 2, K , m

j =1

(2) For numerical attributes, d(·, ·) could attain several metrics. Eq. (3) provides the Minkowski Distance, where p is a positive integer as δ j=0 if the value of j-th attribute is missing, else δ j=1. n

d ( x k , x i ) = P x k − x i P= ( ∑

j =1

p

δ j ( x k j − x ij ) )

1 p

(3)

For nominal attributes, d(·, ·) is defined by the ratio that the length of the Longest Common Substrings (LCS) over the longer of the two strings in eq. (4). s

d ( x k , x i ) = 1 − ∑ length ( LC S j ) m ax_ length ( x k , x i )

(4)

j=0

After

xi is assigned to the arg max U (v j , xi ) -th cluster by the Maximum 1≤ j ≤ K

Subordination

vk =

m

∑ U (v

[16]

Principle

,

the

centroids

are

iteratively

updated

by

m

k

, xi ) * xi

i =1

∑ U (v

k

(5)

, xi )

until

it

meets

i =1

L

K

[17] W ( L , K ) = ∑ ∑ d ( v (jl ) , v (jl −1) ) < ε , where L is the max number of loops . l =1 j =1

37

5

4.2

Outlier Test

Since k-Means is particularly sensitive to outliers, a non-parametric approach of resisting noise is posed to ensure the robustness and correctness during and after the clustering process. Two types of outliers are clarified: a) the furthermost instances away from all the clusters; b) the furthermost instances away from the other majorities within one cluster. The hierarchical strategies, in turn, have been employed. For the first case, these abnormal elements are isolated in the second layer as questionable individual clusters with rarely less than N members when each turn of clustering ends. And the other regular clusters form the first tier. In the next turn of iterative clustering or classifying, the components at the first layer are superior to be considered for reducing computation times. Once an instance lacks distinguishable membership degrees to all at the first layer, the irregular ones at the second layer get involved in queue by members from more to less. For the latter case, disposals are necessary after pre-clustering and the global clustering respectively as follows: one existed cluster might be divided into a big cluster with more members and several non-overlapping small ones with fewer members. 4.3

Estimation and Imputation

Classify each instance of Xic into one of the clusters by eq. (2). And the Entropy Method [19] is utilized for imputation using the info T=(T1,T2,…,Ts)T,s=1,2,…,|ck| for mis

the cluster ck that xi

xijmis =

∈ X ic is attached into. The j-th missing value is estimated by

s

∑

wq t qj (6). The renewed instance is merged into the cluster afterwards

q =1, q ≠ i

with updated centroids.

5

Experimental study

The experiment is to achieve two goals: one is to examine the feasibility of PCFKMI and adjust the parametric values for the optimum impact on the accuracy of imputation; the other is to compare the accuracy that PCFKMI, NNI (Nearest Neighbor Imputation) and KMI (K-Means Imputation) perform under different mechanisms at distinct missing rates with various noise rates. The Root Mean Squared Error (RMSE)

1 m

m

∑

~ ( e i − e i ) 2 is adopted to evaluate the precision of these

i =1

algorithmic methods, where ei is the original value and

~

ei

is the estimated, as m is the ~

total number of imputations. For numerical values, e i − e i is the quantitative difference, yet for nominal ones, only 0 or 1 could be the alternative, that is, 0 for the same value while 1 for textual difference. Lower RMSE implies higher validity. The Abalone [19] dataset is selected as the benchmark in the experiment. It is a multivariate dataset with 8 attributes. One attribute is nominal, and the others are numerical. One

38

6

thousand records from the original dataset are randomly chosen as the population. Particularly, proportional observations in several remaining records according to the missing rate are intentionally set null to simulate the missing data, as some values in specific copies are deliberately substituted by anomalous values at the extent to the coincident noise rate. 5.1

Feasibility

This part of work is conducted under MAR, which is the most usual mechanism.

Fig. 3. RMSE for varying fuzzier

Fig. 4. RMSE for distinct similarity metrics

Fig. 3 illustrates the effect that the variation of fuzzier on RMSE. As a matter of fact, an appropriate fuzzier is crucial to the quality of imputation, and here 1.3 is supposed to be the best. The histogram in Fig. 4 displays how the four kinds of similarity metrics impact RMSE respectively. The Manhattan Distance appears to excel significantly in comparison with other 3 types of distances. Thus, it is considered as the unique similarity metric thereafter.

(a)

(b)

Fig. 5. RMSE (a) & Time (b) for varying pre-clustering rate

The pre-clustering on “small” samples is such an equivocal concept that we attempt to choose less than 50% of the total number to clarify it in the test. Fig. 5 (a) shows the trend that RMSE declines with increasing pre-clustering rate (which refers to the proportion that the amount of instances for pre-clustering compared to the total number of instances). The fact confirms that more initial instances for pre-clustering perform better. In contrast, the time curve in Fig. 5 (b) monotonously rises when preclustering rate increases. So 30% is the compromising point.

39

7

(a)

(b)

Fig. 6. RMSE (a) & Iteration Times (b) for varying iteration times

Fig. 6 (a) demonstrates that RMSE decreases and descends at a threshold as iteration times arise. Fig. 6 (b) describes that 15 times iterations could bring the variation to a fine convergence actually. 5.2

Comparison

Table 1. RMSE for three methods at varying missing rates / noise rates / missing mechanisms Noise Rate Missing 0% 10% Mechanism Rate NNI KMI PCKMI NNI KMI PCKMI NNI 5% 0.325 0.292 0.284 0.314 0.308 0.298 0.332 MCAR 10% 0.318 0.314 0.254 0.337 0.324 0.300 0.353 15% 0.335 0.320 0.304 0.342 0.320 0.262 0.356 5% 0.288 0.282 0.257 0.299 0.298 0.300 0.333 10% 0.285 0.277 0.278 0.282 0.304 0.265 0.292 MAR 15% 0.334 0.300 0.238 0.348 0.311 0.301 0.341 5% 0.242 0.222 0.220 0.262 0.248 0.240 0.256 NMAR 10% 0.269 0.264 0.256 0.269 0.269 0.268 0.283 15% 0.280 0.261 0.242 0.287 0.271 0.250 0.297

20% KMI 0.331 0.333 0.330 0.328 0.325 0.330 0.267 0.279 0.290

From the data in Tab. 1, we can summarize it as follows: firstly, experimental result under NMAR proves to be better for all the three individual algorithmic methods than other two distinct mechanisms, where performance at MAR appreciably precedes that at MCAR; secondly, RMSE always arises when the missing rate increases both globally and respectively for these three approaches; thirdly, PCFKMI maintains

40

8

PCKMI 0.311 0.257 0.310 0.327 0.304 0.315 0.284 0.273 0.283

much better than the other two counterparts and it is capable of properly alleviating the corruption by the noise.

6

Conclusions

The paper studies the missing value treatment mechanism and proposes PCFKMI, an algorithm for missing value imputation based on fuzzy k-Means clustering. The approach makes use of samples for pre-clustering and obtains the optimum entire clustering case by fuzzy statistical evaluation. Appropriate detection and process on noisy data could also guide the rest of instances to more precise groups, which is implicitly helpful to produce more correct estimation on different types of parametric values. The experimental results illustrate that the PCFKMI generally performs well under all the three missingness mechanisms and excels NNI and KMI on validity. Acknowledgments. This work is supported by National Natural Science Foundation of China (61003016) and Project of State Key Laboratory of Software Development Environment, BUAA (SKLSDE-2011ZX-09).

References 1. 2.

3. 4. 5. 6. 7.

8. 9. 10.

11. 12.

Li D, Gu H, Zhang L (2010) A fuzzy c-means clustering algorithm based on nearestneighbor intervals, Expert Syst Appl, vol. 37, no. 10, pp. 6942–6947. Scheel I, Aldrin M, Glad IK, Sorum R, Lyng H, and Frigessi A (2005) The influence of missing value imputation on detection of differentially expressed genes from microarray data. Bioinformatics, vol. 21, no. 23, pp. 4272-4279. Zhang S, Jin Z, Zhu X (2011) Missing data imputation by utilizing information within incomplete instances, J Syst Software, vol. 84, no. 3, pp. 452-459. Allison PD (2001) Missing Data. Dissertation, Sage University Papers Series on Quantitative Applications in the Social Sciences, Thousand Oaks, Sage, CA, pp. 7-136. Little RJA, Rubin DA (2002) Statistical analysis with missing data, 2nd edn. John Wiley and Sons, NY. Altmayer L, Available at: http://analytics.ncsu.edu/sesug/1999/075.pdf. Beaumont JF (2000) On regression imputation in the presence of nonignorable nonresponse. In Proceedings of the Survey Research Methods Section, American Statistical Association, pp. 580-585. Dempster AP, Laird NM, Rubin DB (1977) Maximum likelihood from incomplete data via the EM algorithm. J R Stat Soc: Series B (Methodological), vol. 39, no. 1, pp. 1-38. Qin Y, Zhang S, Zhu X, Zhang J, Zhang C (2007) Semi-parametric optimization for missing data imputation. Appl. Intell., vol. 27, no. 1, pp. 79-88. Liao Z, Lu X, Yang T, Wang H (2009) Missing Data Imputation: A Fuzzy K-means Clustering Algorithm over Sliding Window, In Proc. of The 6th International Conference on Fuzzy System Knowledge Discovery, pp. 133-137. Magnani M (2004) Techniques for Dealing with Missing Data in Knowledge Discovery Tasks, Available at: http://magnanim.web.cs.unibo.it/data/pdf/missingdata.pdf Batista G., Monard M (2003) An analysis of four missing data treatment methods for supervised learning, Appl Artif Intell, vol. 17, no. 5-6, pp. 519-533.

41

9

13. 14.

15. 16. 17.

18. 19.

Zhang S, Zhang J, Zhu X, Qin Y, Zhang C (2008) Missing value imputation based on data clustering, IEEE Transactions on Computational Science I, pp. 128-138. Van Hulse J, T.M. Khoshgoftaar (2008) A comprehensive empirical evaluation of missing value imputation in noisy software measurement data, J Syst Software, vol. 81, no. 5, pp. 691-708. Zhu B, He C, Liatsis P (2012) A robust missing value imputation method for noisy data, Appl Intell, vol. 36, no. 1, pp. 61-74. L.A. Zadeh (1965) Fuzzy sets, Information and Control vol. 8 no. 3, pp. 338–353. Li D, Deogun JS, Spaulding W, Shuart B (2004) Towards Missing Data Imputation: A Study of Fuzzy K-means Clustering Method. In Proc. of the Rough Sets and Current Trends in Computing, pp. 573-579. Chen H, Huang DC (2011) Missing Data Imputation Based on Generalized Mahalanobis Distance, Computer Science, vol. 38, no. 5, pp. 149-153. Waugh S, http://archive.ics.uci.edu/ml/datasets/Abalone

42

10

Web Robot Detection based on Monotonous Behavior Shinil Kwon1, Myeongjin Oh1, Dukyun Kim1, Junsup Lee2, Young-Gab Kim1, and Sungdeok Cha1, ∗ 1

College of Information and Communication, Korea University, Anam-dong Seongbuk-Gu, Seoul, 136-701 South Korea {shinilkwon1, iamohmj, kim9069}@gmail.com, {always, scha}@korea.ac.kr 2 Convergent Contents Research Department, Electronics and Telecommunication Research Institute, 161 Gajeong-dong, Yuseong-gu, Daejeon, 305-700, South Korea [email protected]

Abstract. Several studies examined various features on how to most effectively detect web robots. Based on an insight that most web robots, regardless of specifics, would exhibit focused and therefore monotonous behavior, this paper proposes that monitoring the rate of behavioral change is highly effective in detecting sessions initiated by web robots. Empirical evaluation performed on more than one billion requests made to www.microsoft.com web servers confirms that “switching factor” is indeed highly effective. In this paper, we explain the three features whose switching factor was used in web robot detection. Unlike previous studies where the types of web robots to detect were limited to a few types (e.g., text crawlers, link checkers, and email harvester), we extend the types of web robots to image collectors, icon collectors, download tools, etc. Keywords: web robot detection, classification, data mining, switching factor

1

Introduction

Accurate and rapid detection of various web robots has become an important research topic with the explosive growth of the web-based economy. In addition to text crawlers deployed by search engine companies, there are a variety of web robots such as paper crawlers, link checkers, email harvesters, and icon crawlers. Some hostile web robots pose serious security threats. One must be able to distinguish sessions initiated by interactive users surfing the web from those created by programmed web robots. D. Doran and S. Gokhale [1] classified existing robot detection techniques into four categories: syntactical log analysis, analytical learning techniques, traffic pattern analysis, and Turing test systems. Among these categories, analytical learning techniques and traffic pattern analysis closely relate to the anomaly detection used by this study. The analytical learning techniques focused on finding various features of each web robot (e.g., 26 or 34 features) with a running average or standard deviation, and ∗

Corresponding author

adfa, p. 1, 2012. © Springer-Verlag Berlin Heidelberg 2012

43

finally, concentrate on web robot detection using a data mining algorithm. Tan & Kumar [2] examined web robot detection in terms of 26 features such as the ratio of empty strings in the referrer field and the ratio of the head method with the C4.5 decision tree algorithm. Lourenco and Belo [3] chose roughly same features and C4.5 decision tree algorithm to Tan and Kumar. Bomhardt et al.[4] added eight features and applied a neural network using web log preprocessing tool called. Stassopoulou and Dikaiakos [5] constructed a Bayesian network using six features such as Maximum sustained click rate, Duration of session. Lu and Yu [6] detected web robots using hidden Markov model (HMM) and inter arrival time. The traffic pattern analysis proposed new features based on behavioral analysis of web robots. Almeida et al. [7] presented a characterization of the workload generated by web robots. This characterization focuses on the statistical properties of the arrival process and on the robot behavior graph model. Huntington et al. [8] proposed a multi-step log analysis technique with online scientific journal Glycobiology. Duskin and Feitelson [9] studied the interaction between the query submittal rate and the minimal interval of time between different queries. Ye et al. [10] measured a server workload using difference of HTTP response time and ICMP response time. Web robot detection using a switching factor, as reported in this paper, overcomes limitations of other studies: our approach detected each web robot with equal performance using our new features. Although the six different types of web robots had great imbalance in the number of sessions, we successively detected each web robot without discarding the text crawler sessions by using only three features common to all web robots. The remainder of our paper is organized as follows. In Section 2, we introduce the concept of switching factors, with an illustrative example. In Section 3, we explain the Microsoft web log used in this study. Section 4 details experimental results and compares our results against those reported by Tan and Kumar. Section 5 concludes the paper.

2

Switching factor

Our approach is based on an insight that all web robots, regardless of their specific tasks they have been programmed to perform, would exhibit highly focused and therefore consistent behavior unlike that exhibited by interactive users. Web robots may employ diverse search strategies (e.g., depth-first vs. breath-first), algorithms, or data structures. For example, we discovered in Microsoft log that some sessions initiated by an image crawler seldom contained image requests. While strange at first, further analysis revealed that image requests were made by different and spawned threads. Different image crawlers left traces indicating that crawling and image requests were all made in the same thread. Because such image crawlers exhibited vastly different in strategies, direct comparison of features may yield little similarities. However, they exhibited highly focused and consistent behavior, and “switching factor” for both web robots turned out to be extremely small when compared to the web sessions generated by interactive users.

44

Fig. 1. Is there a request to an image in the log?

Figure 1 illustrates the switching factor concept. It illustrates a session in which image requests occur in the first, fifth, and sixth requests. Rather than counting the ratio of image requests (e.g., 3 out of 6 in this example), as [2] did, we observe that “behavioral changes” occurred at the second and fifth requests. For a session with N requests, there can be at most (N-1) opportunities for changes. The switching factor in this example is 0.4 (or 2 out of 5) as defined below: Definition 1 For each feature F, switching factor (SF) is the ratio of changes between all the adjacent requests.
 

Number of requests whose F changed from the preceding request Number of requests  1

3

A proposed approach to web robot detection

3.1

Web server log www.microsoft.com

The web server log used in this study is same as that used in [11]. It is about 250 GB. More than one billion requests are captured in 24 hours. We divided the web log into four segments, each containing six hours, to empirically evaluate web robot detection accuracy. Care was taken to ensure that “peak” hours in different continents (e.g., North America, Asia, and Europe) are properly reflected in each segment. Data belonging to each segment is further divided into training and test data in a 2:1 ratio. That is, logs collected during the first four hours were used as training data, and the rest were used as test data. 3.2

Selection of detection features

We selected three fields in the IIS format log as initial candidates based on an earlier study [11], built databases containing sessions initiated by interactive users as well as web robots, computed switching factors on each features. Switching factor on unassigned referrer field (cs-referrer). Conventional wisdom on web robots is that referrer field would contain empty string (e.g., unassigned). While generally true, our analysis on Microsoft log revealed exceptions. Girafa and Arachmo had string value included in the referrer field in more than 80% of the requests. All other web robots almost always had unassigned referrer

45

field. That is why inclusion of empty string is not always an effective, though straightforward, feature. However, switching factors are consistently Switching factor of file types (cs-uri-stem). While Microsoft logs consist of various types of file such as web pages or video files, almost all their requests are web page and image files. (e.g., Web page: 70.5%, Image file: 29%) Interestingly, the requesting patterns in these two file types are entirely different. With the apparent exception of image or icon crawlers, web robots are less likely to be interested in images. Text crawlers and link checkers are the most obvious samples. Among the text crawlers, types of requested resource varied depending on who launched them (e.g., Google vs. Yahoo). For interactive users who generally traverse pages by clicking the links, actual request would be generated by a browser. In such cases, image requests would be mixed. In fact, interactive users rarely requested only images (i.e., 1.9%), and such superficial criteria would be ineffective in separating behavior of text crawlers from that of interactive users. However, switching factors on web robots are unanimously and significantly below that of interactive users. One must ask whether or not behavioral patterns change frequently between the adjacent requests because switching factor analysis on web robots returned unanimously and significantly lower values compared to that of interactive users. Switching factor on number of bytes from clients to the server (cs-bytes). Analysis on the number of bytes clients sent to the server revealed similar pattern to other features. Lee et al. [11] reported that majority of web robots, except the image crawler named Girafa, sent smaller number of bytes (e.g., 320 bytes) than interactive users did (e.g., 490 bytes) in almost all requests (e.g., 95.57% to 100%). About 8% of Girafa requests, however, did not follow the pattern. While generally useful, such sporadic exceptions may result in false alarms in web robot detection. We instead compare the current request size against the running average of all the requests and ask if there are changes in patterns (e.g., below or above the average). Such analysis resulted in consistent and substantially lower values when applied on web robots. This is yet another evidence of highly monotonous behavior of web robots.

4

Experimental evaluation

F-measure, computed from recall and precision values, is widely used in machine learning research. It provides an impartial viewpoint in that one can increase the value of one at the expense of the other. On one extreme, an algorithm may achieve 100% recall value if it simply declares all the sessions as those initiated by web robots. Such algorithm is obviously useless and filled with numerous false alarms. In running the experiment, we used an open-source implementation of machine learning algorithm commonly known as C4.5 [12]. Many research groups working on web robot detection, including Tan and Kumar, used the same software. We changed session lengths (e.g., the minimum number of requests included in a valid session)

46

from three to eight, repeatedly measured precision and recall values for the chosen features. For example, if session length is set to eight, all the sessions containing seven or less requests are excluded from the training and test data. a

Text crawler

Proposed approach Tan & Kumar 100

90

90

80

80

Accuracy (F-measure)

Accuracy (F-measure)

100

70

60

50

Link checker

b

70

60

50

40

40

30

30 3

4

5

6

7

3

8

4

Email collector

Proposed approach Tan & Kumar 100

90

90

80

80

Accuracy (F-measure)

Accuracy (F-measure)

c

70

60

50

40

6

7

8

d

Image collector

Proposed approach Tan & Kumar

70

60

50

40

30

30 3

4

5

6

7

8

3

4

# of requests

e

5

6

7

8

# of requests

Icon collector

100

Proposed approach Tan & Kumar 100

90

90

80

80

Accuracy (F-measure)

Accuracy (F-measure)

5

# of requests

# of requests

100

Proposed approach Tan & Kumar

70

60

50

f

Download tool

Proposed approach Tan & Kumar

70

60

50

40

40

30

30 3

4

5

6

7

3

8

4

5

6

7

8

# of requests

# of requests

Fig. 2. F-measures grouped by web robot types on various session lengths

As already mentioned, the common features, such as switching factor, are one solution to imbalance problems. Although we have imbalanced samples, like other studies [2-6], we successively classified six types of web robots with similar performance. The decision tree using the proposed features achieved about 91% accuracy in a session composed of more than six requests. The download tool achieved relatively low performance (about 88% accuracy). This results in a greater switching factor in terms of packet size from client than other types of web robots, since download tools collect various types of files. This finally leads to low performance.

47

We were able to use 16 out of 26 features, in replicating an experiment reported in [2], due to differences in web log format. Our approach and that of Tan are both interested in text crawler, email collector, and link checker. Our approach in detecting these three types of web robots is slightly more accurate when sessions exceed six requests.

5

Conclusions and future works

This paper demonstrated that behavioral patterns are different between the sessions initiated by interactive users are fundamentally different from those triggered by web robots. Switching factors, characterizing the degree of uniformity in request patterns, have been proposed as effective features to detect web robots of various types. Proposed features works well on all types of major web robots found in the Microsoft web server log. Real-time detection of web robots seems feasible because proposed features are relatively simple to compute and small in numbers. Real world demonstration remains a major task for future research.

References 1. Doran, D., Gokhale, S.: Web robot detection techniques: overview and limitations. Data Mining and Knowledge Discovery 22(1),183-210 (2011) 2. Tan, P.N., Kumar, V.: Discovery of Web robot sessions based on their navigational patterns. Data Mining and Knowledge Discovery 6(1), 9–35 (2002) 3. Lourenco, A., Belo, O.: Catching web crawlers in the act. In: Proc. of the 6th international Conference on Web Engineering, ACM, pp. 265-272 (2006) 4. Bomhardt, C., Gaul, W., Schmidt-Thieme, L.: Web robot detection-preprocessing web logfiles for robot detection. New developments in classification and data analysis pp. 113-124 (2005) 5. Stassopoulou, A., Dikaiakos, M.D.: Web robot detection: A probabilistic reasoning approach. Computer Networks 53(3), 265–278 (2009) 6. Lu, W., Yu, S.: Web robot detection based on hidden markov model. In: Communications, Circuits and Systems Proc., 2006 International Conference on, IEEE, vol. 3, pp. 1806-1810 (2006) 7. Almeida, V., Menasc´e, D., Riedi, R., Peligrinelli, F., Fonseca, R., Meira Jr., W.: Analyzing Web robots and their impact on caching. In: Proc. of the Sixth Web Caching and Content Delivery Workshop (2001) 8. Huntington, P., Nicholas, D., Jamali, H.: Web robot detection in the scholarly information environment. Journal of Information Science 34(5), 726-741 (2008) 9. Duskin, O., Feitelson, D.G.: Distinguishing humans from robots in web search logs: preliminary results using query rates and intervals. In: Proc. of the Workshop on Web Search Click Data, pp. 15–19. ACM (2009) 10. Ye, S., Lu, G., Li, X.: Workload-aware web crawling and server workload detection. In: Proc. of Asia Pacific Advanced Network 2004 11. Lee, J., Cha, S., Lee, D., Lee, H.: Classification of web robots: An empirical study based on over one billion requests. Computers & Security 28(8), 795–802 (2009) 12. Quinlan, J.: C4. 5: programs for machine learning. Morgan kaufmann (1993)

48

A Configuration Management Model on the High-speed Networks Jie Huang, Lin Chen 1

School of Computer, National University of Defense Technology, Chang Sha, China [email protected]

Abstract. The infiniband-based high-performance cluster system can not only provide the large bandwidth and the low latency, but also achieve CPU bypass operations during the data exchange. Being the part of the IBA (Infiniband Architecture), the management model runs through each layer, and brings some important RAS (Reliability, Availability and Scalability) mechanisms to IBA. The IBA network configuration management, which is the most important part of the IBA management model, is the basic interconnection technique of the infiniband network. In this paper, the IBA network configuration management has been discussed. We propose a network configuration management model in the infiniband network named NCM-IB. Based on that, we describe the selection mechanism of master NCM, the topology discovery mechanism, the routing algorithm and the distribution mechanism. The NCM-IB model can effectively maintain the normal operation of the infiniband network in the practical application. Keywords: Infiniband Architecture; Network Configuration Management; Topology Discovery; Distribution Mechanism

1

Introduction

Infiniband-based high-performance cluster system can not only provide large bandwidth and low latency, but also achieve CPU bypass operations during data exchange[1,2]. It can also implement zero-copy based on infiniband, which can maximize the CPU performance and improve the overall performance of cluster systems significantly[3]. Being the part of IBA (Infiniband Architecture), management model runs through each layer, and brings some of important RAS (Reliability, Availability and Scalability) mechanisms to IBA[4,5]. The features of IBA management model, which include topology discovery, configuration, communication and fault-tolerant and so on, support the management components of the multi-suppliers and the network upgrade, which ensures both the interoperability of different version products from various suppliers and the integration with enterprise-level management tools of data center[6-8]. IBA network configuration management, which is the most important part of IBA management model, is the basic interconnection technique of infiniband.

49

In this paper, IBA network configuration management has been discussed. We propose the network configuration management model in infiniband network named NCM-IB. Based on that, we present the selection mechanism of the master network configuration manager (NCM), the topology discovery mechanism, and the up/down routing algorithm and the distribution mechanism. The NCM-IB model can effectively maintain the normal operation of the infiniband network in practical application.

2

The Configuration Management Model

The infiniband networks need the correct configuration before the normal communication. The NCM-IB model is built on top of user layer verbs source language and user layer MAD (Management Datagram), which transmits information and configures network through HCA (Host Channel Adaptor) hardware and the network [9-11]. By following the object-oriented framework, The NCM-IB model denotes physical component of infiniband network as the class in object-oriented program, and the relationship between classes reflects the association between physical components which include nodes, switches, ports, network configuration manager and so on. For example, each node object contains several port objects, and each port object point to another port object, which belongs to some switch object. All device objects are organized according to certain relationship in the infiniband network, which is composed of the infiniband network objects, the NCM objects, SA objects and some other auxiliary objects.

3

The Management Mechanism

The NCM-IB is built on top of user layer verbs source language and user layer MAD, which transmits information through some base hardware including HCA, switches in infiniband network and the other hosts. Basing on the NCM-IB model, we describe the selection mechanism of the master NCM, the topology discovery method, the routing algorithm and the distribution mechanism.

3.1

The selection mechanism of the master NCM

The NCM runs on the port of HCA. However, if there are more than one NCM in the network, only one of them can be the master NCM, and the rest of them will become backup NCM. Meanwhile, each port can only have one NCM. The selection of the master NCM is a part of the initialization procedure, and is the key factors of the successful initialization and the configuration in the infiniband network. The main effect of master NCM is as follows: 1) Discovery of physical topology in the current network. 2) Allocation of local identifiers (LID) for each node, switch and router.

50

3) To determine a reasonable path between each pair of nodes. 4) To scan the network, and discover the change of network topology, then handle node joining and leaving. In the NCM model, there are four states including discovering, standby, master and not active. In the initialization of the infiniband network, the NCM node will turn itself into discovering state, and begin to scan the network. While discovering a prior NCM or the master NCM in the network, the current NCM node will turn itself into standby state and inquiry the master NCM continuing from time to time, and if the master NCM do not respond, which means it has expired, the current NCM node turn itself back to the discovering state. If the current NCM node hasn’t found any prior or the master NCM at the end of its discovering state, it will turn itself into the master state, and begin to initialize the network. While being the standby state, a NCM isn’t going to inquiry the master NCM, it can turn itself into not active state. The master NCM scans the network at the predetermined time, and if it discovers a prior NCM which is in the standby state, it will transfer its master identity to this prior NCM by sending management message, and turn itself into the standby state. 3.2

The Topology Discovery Method

After being chosen as the master, the NCM has to scan the topology of the network in order to find out the connection relations of the network, and then allocate the local identifiers (LID) which will be used in the network communications for each node. The topology scan will be executed once during the initialization of the infiniband network, and will be executed according to the predetermined time or the time-trap after that, in order to detect the change of the network topology in time. The NCM scan the network by utilizing MAD in the infiniband network. The MAD transmit message using the direct routing, which ensure that MAD can begin to work before network configuration. In the NCM-IB model, we introduce the network scan and the trap mechanism to detect network changes, in order to get the latest network topology. We combine the advantage of both complete discovery mechanism, which discards previously configured information in network discovery process and discover the latest topology information starting from scratch, and partial discovery mechanism, which scans only the area affected by topology change rather than the whole network, in order to reduce the number of MAD and computational overhead during network topology discovery. During completing the discovery mechanism, the NCM object discards all collected information and starts to discover the complete network topology while being aware of the change of network topology. The discovery procedure has to use direct routing, which is relatively slower than LID routing management packets. In order to discover all the active network nodes, the NCM constantly sends corresponding MAD, which will be received by the destination node of each direct route, and induces the corresponding response. For the sake of simple, the topology discovery procedure is considered as spreading in order, which means the spread of scanning MAD is in an uncertain way. The NCM sends a new scanning MAD while receiving respond from other devices of the network. The algorithm of the complete discovery mechanism is shown in Fig.1, including not only respond handling MAD

51

and inquiry sending the MAD in scanning procedure, but also scan handling the MAD in the discovery procedure. When the topology change, including adding or deleting nodes or links, affects only a small part of network, we can activate the partial discovery mechanism which scans only the affected part of network. Because the topology information collected before is fully utilized in the partial discovery mechanism, the number of MAD and the computational overhead during topology discovery can be greatly reduced.

Fig. 1. The description of the discovery algorithm.

In this paper, we implement the partial discovery mechanism with both direct routing and LID routing, in order to reduce the transmission overhead. While a topology change occurs, the NCM can detect at least one switch port which has changed its state during the periodic network scanning. If a switch port becomes active, the NCM begins to discover new devices on the other side of this port, and scan unknown area as complete discovery mechanism. Scanning MAD can make use of a section of initial LID routing to reach a node, which fully utilize the information that already exists in forwarding table. We can reach a switch that has just changed its state by initial LID routing, but the path from the changing switch port to new device must be tracked by direct routing. That is because there is no useful configuration information in the initialization of network, and scanning MAD has to use direct routing. If a switch port stops activity in changing list of the NCM, a part of the network will be no longer reachable. We put all unreachable devices into a collection U (Unreachable), and search all ports of reachable devices that directly connect to U according to the existing network topology information. By using these reachable ports, we can scan devices in U by direct routing, and denote devices that can not be scanned successfully as M (Missing).

52

3.3

The Routing Algorithm

The NCM assumes that the network topology is irregular, especially in the large scale network. It is easy to design a scalable, flexible routing system on the irregular network topology, however, the irregularity of network topology also makes packets routing and the deadlock avoidance mechanism too complicated. The Each link in the network will be denoted as either up or down direction in up/down routing algorithm. The turn of message transmission from the down direction to the up direction will be forbidden, which avoids the circular dependency between the channels, and make the algorithm deadlock-free. The routing algorithm is described as follows: 1) Select the node which has the smallest LID as the root node, and build a BFS (Breadth-First Search) spanning tree. 2) Assign the direction for each link. Considering a link that connects two switches, one direction of the link has been denoted as the up direction, and the other direction will be denoted as the down direction. The up direction of a link is defined as the closer direction to the root node of the BFS spanning tree. 3) If the both switches the link connects located in the same level of spanning tree, we will take the side of switch that has smaller LID. The messages from source node firstly go along with zero or more up direction links, and then go along with zero or more down direction links, and finally get to the destination node. From the routing strategy of the routing algorithm, we can see that neither ‘up direction’ links nor ‘down direction’ links can form a routing loop. Each node in the spanning tree will eventually reach the root node by going along with the up direction links. The root node can reach any node in the spanning tree by going along with the down direction links. Therefore, for any pair of nodes in the tree, a node firstly go along with zero or more up direction links, and then go along with zero or more down direction links, and finally it will get to any other node. A path that connects any pair of nodes can be formed in this way. In the path that messages pass, it could prevent from the deadlock happens that the turn of message transmission from the down direction to the up direction is forbidden. If the destination node is further from the root node or the LID of the destination node is larger, a node will choose the down direction links to forward its messages. If the destination node is closer from the root node or the LID of the destination node is smaller, a node will choose up direction links to forward its messages, and then its messages will go along with some down direction links to reach the destination node. 3.4 The Distribution Mechanism Basing on the topology discovery method and the routing algorithm described above, we can construct a forwarding table for each switch in the network. In the forwarding table, the address field stands as the destination LID, and the port field stands for the port to be used. The forwarding table in the switch must include all legal LID in order to forward all packets correctly, so we have to generate one table item in the form of for any legal LID when constructing the forwarding table for each switch. Considering each legal LID in forwarding table for each switch, if this LID

53

stands for a switch, the forwarding port is corresponding port field value of the minimum LID in the LID matrix. If the LID stands for a HCA, we firstly find the LID of the switch that directly connects to this HCA, and the forwarding port is corresponding port field value of the minimum LID in the LID matrix. Based on the forwarding table which has been constructed successfully, the management MAD of the NCM will configure the forwarding table to corresponding switches and activate the HCA which connects to the switch port, and finish the configuration of the infiniband network

5

Conclusion

In the paper, we propose the IBA network configuration management model named the NCM-IB model. Based on that, we describe the selection mechanism of the master NCM, the topology discovery mechanism, the routing algorithm and the distribution mechanism. The NCM-IB model can effectively maintain the normal operation of the infiniband network in practical application. Acknowledgement. This work is supported by Program for Changjiang Scholars and Innovative Research Team in University (No.IRT1012), "Network technology" Aid program for Science and Technology Innovative Research Team in Higher Educational Institutions of Hunan Province and Hunan Provincial Natural Science Foundation of China(11JJ7003).

References 1. International Business Machines (IBM) Corporation, “InfiniBlue Host Channel Adapter Access Application Programming Interfaces Programmer's Guide,” October 2002. 2. Mellanox Technologies Inc., “Mellanox IB-Verbs API (VAPI),” 2001. 3. Ron Brightwell and Arthur B. Maccabe, “Scalability Limitations of VIA-Based Technologies in supporting MPI,” Proceedings of Fourth MPI Developer's and User's Conference, March 2000. 4. Ron Brightwell, Arthur B. Maccabe, and Rolf Riesen, “The Portals 3.2 Message Passing Interface Revision 1.1,” Sandia National Laboratories, November 2002. 5. Vieo Inc., “Channel Abstraction Layer (CAL) API Reference Manual V2.0,” January 2002. 6. Intel Corporation, “Linux InfiniBand Project,” http://InfiniBand.sourceforge.net. 7. InfiniBand TM Architecture Specification Release 1. 2. 1 [ EB/OL ] . [ 200922218 ] . http :/ / www. infinibandta. org/ . 8. Bermúdez, Casado R, Quiles F J,et al. Evaluation of a Subnet Man-agement Mechanism for Infiniband Networks [C]//Proc of Int'l Conf on Parallel Processing, 2003:117-124. 9. Ranjit Noronha, Dhabaleswar. Designing High Performance DSM Systems Using InfiniBand Features.http://nowlab.cis.ohiostate.edu/publications/conf-papers/2004/noronha-cgrid04.pdf, 2004-05. 10. Jiesheng Wu, Pete Wyckoff, Dhabaleswar K.Panda.PVFS over InfiniBand: Design and Performance Evaluation. http://nowlab.cis.ohio-state.edu/publications/confpapers/2003/wuj-icpp03.pdf, 2003-07. 11. IBTA.2000.Supplement to InfiniBand TM Architecture Specification, Volume 1. http://www.infinibandta.org/, 2004-12.

54

Challenges and Solutions of Cyber-Physical Systems Lichen Zhang, Jifeng He and Wensheng Yu Shanghai Key Laboratory of Trustworthy Computing, East China Normal University Shanghai 200062, China [email protected]

Abstract. Cyber physical systems(CPS) include a lot of high complexity computing such as dynamic analysis and verification of continuous dynamic property , analysis and verification of real-time property, analysis and verification of spatial property, scheduling and fault tolerance. The correctness of computing results of cyber physical systems not only meets the time requirements, but also meets the spatial requirement, this make that it is impossible to solve the several difficult computing problems by traditional methods. In this paper, we describe some of the research directions that we are taking toward addressing some of the challenges involved in building cyber physical systems. The ultimate goal is to develop foundations and techniques for building safe and effective cyber physical systems. Keywords: Cyber Physical Systems, continuous dynamic, temporal-spatial requirements, non-functional requirements

1

Introduction

Cyber physical systems [1] are dynamical systems with interacting continuous-time dynamics and discrete-event dynamics. Hybrid behavior arises in many contexts, both in man-made systems and in nature. Continuous systems such as walking robots, flight systems, train systems and process control systems, are well-suited to be modeled as cyber physical systems. In these examples, the continuous dynamics model system motion or chemical reactions, while the discrete dynamics model the sequence of contacts or collisions in the gait cycle, cell divisions, valves and pumps switching, control modes, and coordination protocols. The system dynamics are complex enough that traditional analysis and control methods based solely on differential equations are not computationally feasible. To understand the behavior of cyber physical systems, to specify, to analyze, to design, to verify, and to control these systems, theoretical advances and analytical tools are needed. In this paper, we describe some of the research directions that we are taking toward addressing some of the challenges involved in building cyber physical systems. The ultimate goal is to develop foundations and techniques for building safe and effective cyber physical systems.

55

2

Research Challenges of Cyber Physical Systems

The complexity of future cyber-physical systems is beyond the state of current theory and practice in reasoning about time and space; dynamic analysis and verification of continuous property; middleware for cyber physical systems; temporal-spatial scheduling; non-functional requirement analysis and design; security; fault tolerance; and even systems analysis and development. Space and Time: Cyber physical systems are spatio-temporal [2] in the sense that correct behavior will be defined in terms of both space and time. As cyber physical systems, computations must be completed within established response times, for which we will need new notions of timeliness [3], but they may also have varying temporal requirements based on the current frame of time. As spatial systems, the computations performed and their timeliness will be dependent on spatial behaviors of cyber physical systems. Complex Continuous Dynamics: The dynamic behavior of cyber physical systems is captured in hybrid models[4]. Frequently in hybrid systems in the past, the event-driven dynamics were studied separately from the time-driven dynamics, the former via automata or Petri net models and the latter via differential or difference equations. To understand fully the system’s behavior and meet high performance specifications, one needs to model all dynamics together with their interactions. Reachability questions for systems with complex continuous dynamics are among the most challenging problems in verifying cyber physical systems. Middleware for Cyber Physical Systems: Several challenging issues for cyber physical systems such as communication delays, the interface between a control application and the network layer, platform heterogeneity, clock differences between the physical layer, communication layer and computation layer, arise. One solution to release these burdens from the software engineer is to interpose an abstraction layer between the physical layer and the underlying networked communication and computation layer. Typically, such an abstraction can be realized as a software framework, called a middleware [5]. Existing middleware systems lack the flexibility needed to support cyber-physical systems with diverse application semantics and requirements. Therefore, a key open challenge for distributed cyber physical systems is to develop a flexible middleware infrastructure that can be easily configured to support the diverse requirements of cyber-physical systems. Temporal-Spatial Scheduling: Designing CPS is a challenging task since it has to consider spatial requirement, in particular, if a task is bound to a specific physical location, this requires a coordination in space and time. However, the current real-time task models and supporting scheduling theory make time a “first class” property. New models that integrate space and time will be much better suited for cyber physical systems [6]. Non-Functional Requirements: Non-functional requirements address important issues of quality and restrictions for cyber physical systems, although some of their particular characteristics make their specification and analysis difficult. The non-functional properties are highly important for cyber physical systems as they are designed to operate in environments where failure to meet these non-functional requirements can have enormous cost both from financial, influential or physical

56

aspects. Therefore it is essential that these properties are calculated as precisely as possible during the design and operation of such systems. Security: Security is very important for many CPS applications, such as military target tracking and security monitoring. Providing security to the sensor nodes is challenging, due to the limited capabilities of sensor nodes in terms of computation, communication, and energy. Due to resource limitations as well as vulnerability to physical capture of nodes, traditional public key security protocols are too complicated and energy-consuming for large-scale wireless sensor networks. The privacy and security issues posed by cyber physical still represent a rich field of research problems [7]. Fault-Tolerance: Fault tolerance [8] is usually applied by means of redundancy and diversity. Redundant hardware involves extra software coordination, which makes the software system more complex and prone to errors. Software fault tolerance may be implemented by software re-execution or multiple versions techniques, which also requires the application of additional control mechanisms. Generally a cyber physical system executes a series of tasks subjected to deadlines and jitter constraints. For many of these applications, there are serious limitations in physical size and energy consumption, which imply reduced processing power and memory size. Thus the fault tolerance for cyber physical systems poses a great challenge. Analysis and Design Methods: Currently, there are few tools for modeling and analyzing large-scale deeply embedded systems to guide the designers. However, these tools are isolated efforts which cater to specific problems in designing the system or are geared for certain applications. Therefore, there is a tremendous need for a complete design suite for deeply embedded CPSs, which will be used for modeling, analysis and validation throughout the lifecycle of the system. This suite will provide facilities to model any application for deeply embedded systems and estimate the system performance using mathematical analyses. Moreover, it will address all the major issues that are crucial for the designers, such as lifetime, real-time, security, coverage and reliability.

3

Solutions to Cyber Physical Systems Research Challenges

A number of the methodology and tools have been proposed for the design of cyber physical systems. However, they are still evolving because systems and software engineers do not yet completely understand the cyber physical system requirements, and the complex dynamic behavior of cyber physical systems is difficult of capture in a simple understandable representation. In our opinion, an acceptable cyber physical system design methodology must synthesize the different views of systems, use a number of different methods, and consist of an orderly series of steps to assure that all aspects of the requirements and design have been considered. The timing analysis: In timing analysis phase, the specification of cyber physical systems is first extended with time annotation to express with the timing requirements, that is the specification is annotated with task parameters: worst-case execution time,

57

deadline and period, etc. Then specification is mapped to the precedence graph according to mapping rules. Task parameters together with the precedence graph are used to perform timing analysis, that is, allocation and schedulability analysis according to scheduling algorithms. We believe that you must demonstrate that the requirements are met at two levels of detail and for each class of task. At the macroscopic (system-wide) level we need to show first that all critical tasks will always meet their deadline and that all noncritical tasks satisfy overall requirements. Alternatively, the requirements might be to optimize the value of noncritical hard and soft real-time tasks executed by the system. At the microscopic level, we also would like some level of predictability. For the critical tasks we already know that they will always meet their deadline. For the other real-time tasks their performance will depend on the current state of the cyber physical system. Spatio-temporal reasoning: The main aim of spatio-temporal research is to introduce a hierarchy of languages intended for qualitative spatio-temporal representation and reasoning, provide these languages with topological temporal semantics construct effective reasoning algorithms and estimate their computational complexity. There exist a number of qualitative constraint calculi that are used to represent and reason about temporal or spatial configurations. However, there are only very few approaches aiming to create a spatio-temporal constraint calculus. We use the spatial calculus RCC-8 and Allen’s interval calculus [9] in order to construct a qualitative spatio-temporal calculus. When adding the restriction that the size of the spatial regions persists over time, or that changes are continuous, the calculus becomes more useful, but the satisfiability problem appears to be much harder. Continuous Dynamics: The standard approach to dealing with continuous dynamics for cyber physical systems is to use symbolic or numerical solutions of their respective differential equations. Unfortunately, the range of systems that is amenable to these techniques is fairly limited, because even solutions of simple linear differential equations quickly fall into undecidable classes of arithmetic. As a means for verifying cyber physical systems with challenging continuous dynamics without having to solve their differential equations, A solution is to complement discrete induction for loops and repetitions with a new form of differential induction for differential equations. Differential induction is a natural induction technique for differential equations. It is based on the local dynamics of the (right-hand side of the) differential equations themselves and does not need closed-form solutions for the differential equations. To further increase the verification power, differential strengthening or differential cuts are added as a powerful proof technique for refining the system dynamics with auxiliary invariants that can simplify the proof of the original property significantly. The basic insight is that auxiliary properties that are provable invariants of the dynamics can help prove the original property even if it was not provable before[4]. Middleware: One of the main features of cyber physical systems is their need to be dynamically reconfigured in order to properly adapt to the continuous changes in their environment. An appropriate solution to provide reconfigurability is Aspect-Oriented Software Development, which allows that optional functionalities can be enabled/disabled and services can be replaced with lighter implementations which are less resource consuming. So many non-functional requirements like time

58

requirements, spatial requirement, safety, security, privacy, fault tolerance or mobility can be addressed. Temporal-Spatial Scheduling: Spatial requirements need to be guaranteed in cyber physical systems such as transport control systems and flight control systems in addition to timing constraints. Unfortunately, most conventional scheduling algorithms only take one dimension of them into account. We propose a scheduler model that can be used for multi-dimensional scheduling according spatio-temporal constraints. Based on the scheduler model, we propose a heuristic multi-dimensional scheduling strategy, consisting of two steps. The first step is to select one scheduling algorithm according to spatial requirements. In step 2, we use the selected select one scheduling algorithm to scheduling tasks in cyber physical systems according to timing constraints. Non-Functional Requirements: Aspect-oriented software development methods make up object-oriented software development methods in system development needs of non-functional characteristics of the existing limitations question problem. Use separate technology of concerns separates all the crosscutting concerns of the system, and then analyzed, designed, modeled for each cross-cutting concerns, to address crosscutting concerns in object-oriented software development, the code tangling and scattering problems, enhancing the system's modular degree, lowering coupling between modules[10]. Security: In order to resolve the security problem in cyber physical systems，an efficient privacy-preserving scheme is proposed for CPS message authentication， which is based on the bilinear pairing technique for large—scale group communication and synthetically considers the actual situation of the cyber physical systems according to digital signature principles．In the scheme，sensors are classified into private ones and public ones．The messages from private sensors are verified by group signature ， and those from public sensors by identity-based signature．Furthermore，batch technique is used to accelerate the verification of accumulated messages，and quantitative comparison and determination is employed to handle the conflicting information. Fault Tolerance: Today a number of techniques are available when implementing fault tolerance. A new approach to software implementation is aspect oriented programming (AOP) . AOP allows the programmer to implement fault tolerance functionality separately from the functional code and to combine these implementations to form the final system. Analysis and Design Methods: We propose a design methodology of cyber physical systems that is different from current methodologies and serves to help manage the complexity of cyber physical systems. One of the distinguishing aspects of this methodology is in its ability to express timing constraints and spatial constraints and verify if such constraints are met. A second distinguishing aspect of this methodology is that it can decompose the life cycle into ten stages based on the Model based development according to seven views of systems. The ten stages are: State the problem, model physical environment, make functional analysis, make non-functional analysis, make behavior analysis, make timing analysis and spatial analysis, derive a control algorithm, select models of computation, integrate different

59

models ensemble, verify and test. Seven complementary views are: environment, function, behavior, non-functional aspects, time, space, hardware.

4

Conclusion

In this paper, we described some of the research directions that we are taking toward addressing some of the challenges involved in building cyber physical systems. The ultimate goal is to develop foundations and techniques for building safe and effective cyber physical systems. The further work is devoted to develop the analysis, design and verification methods for cyber physical systems. Acknowledgments. This work is supported by national high technology research and development program of China (No.2011AA010101), national basic research program of China (No.2011CB302904), the national science foundation of China under grant No.61173046, No.61021004, No.61061130541), doctoral program foundation of institutions of higher education of China (No. 200802690018), national; science foundation of Guangdong province under grant No.S2011010004905.

References 1.

E. A. Lee and S. A. Seshia, Introduction to Embedded Systems – A Cyber-Physical Systems Approach, Berkeley, CA: LeeSeshia.org, 2011 2. A. Gerevini and B. Nebel. Qualitative Spatio-Temporal Reasoning with RCC-8 and Allen’s Interval Calculus: Computational Complexity. Technical Report. DEA – University of Brescia. May 2002. 3. E. A. Lee, "Computing Needs Time," Communication of the ACM, vol. 52, no. 5, 2009 4. A. Platzer. Logical Analysis of Hybrid Systems: Proving Theorems for Complex Dynamics. Springer, 2010. 426 p. ISBN 978-3-642-14508 5. Akshay Dabholkar, Aniruddha S. Gokhale: An Approach to Middleware Specialization for Cyber Physical Systems. ICDCS Workshops 2009: 73-79. 6. Tan Ying,Vuran M C,Goddard S,et al.A Concept Lattice-based Event Model for Cyber-Physical Systems[C]//Proceedings of the 1 st ACM/IEEE International Conference of Cyber Physical Systems.2010:50-60 7. Ayan Banerjee et al.Ensuring Safety, Security, and Sustainability of Mission-Critical Cyber–Physical Systems, Proceedings of the IEEE, Vol. 100, No. 1, 283 - 299 , 2012 8. R. Alexandersson and P. Ohman. Implementing Fault Tolerance Using Aspect Oriented Programming. LNCS Dependable Computing, vol. 4746/2007, Springer-Verlag, 2007. 9. D. A. Randell, Z. Cui, and A. G. Cohn. A spatial logic based on regions and connection. Proc. KR-92, 165–176, 1992. Morgan Kaufmann. 10. Wehrmeister, M.A., Freitas, E.P., Pereira, C.E., et al.: An Aspect-Oriented Approach for Dealing with Non-Functional Requirements in a Model-Driven Development of Distributed Embedded Real-Time Systems. In: 10th IEEE International Symposium on Object and Component-Oriented Real-Time Distributed Computing,Santorini Island,Greece, May 7-9,pp. 428-432. IEEE Computer Society, Los Alamitos (2007)

60

Red Tide Image Recognition using Roundness and Entropy based on Semantic Feature Sun Park1, Yoo Kang Ji2, ByungRae Cha3 1

Institute Research of Information Science and Engineering, Mokpo National University, South Korea 2 Department of Information and Communication Engineering, DongShin University, South Korea 3 SCENT Center, GIST, South Korea 1 [email protected], [email protected], [email protected]

Abstract. Red tide is a natural phenomenon that changes sea color by harmful algal bloom. There has been a lot of study on red tide due to increasing of red tide damage. However, to automatically recognize the red tide algae images present in previous studies are not enough. It is difficult to recognize red tide algae images because they do not have matching features for recognizing various species of algae. Besides, the accuracy of algae image recognition is low since previous red tide recognition method mostly uses a few species of red tide harmful algae images for training. In order to resolve the above limitation, this paper proposes the red tide algae recognition method using roundness and entropy based on semantic features of LSA (latent semantic analysis). The experimental results demonstrate that the proposed method achieves better performance than other red tide recognition methods. Keywords: red tide algae, image recognition, LSA (latent semantic analysis), roundness, entropy.

1

Introduction

Red tide is a temporary natural phenomenon involving harmful algal blooms (HABs) in company with a changing sea color from normal to red or reddish brown, and which has a bad influence on coast environments and sea ecosystems [1]. The characteristic of red tide blooms in South Korea spreads fast by currents. Since the red tide has a broad occurrence range (e.g., a varied events area as the southern, western, and eastern part of sea) and spreads quickly, the direct detection of red tide blooms in real time for mitigation activity is difficult. Seven of the 63 species of harmful red tide algae damage fisheries in coastal areas of South Korea. The HABs have inflicted massive mortality on fin fish and shellfish, damaging the economies of fisheries for almost every year from 1993 to 2007. Particularly, there were huge harmful algal blooms of cochlodinium polykrikoides in 1995, resulting in about 95 million dollars in fishery damage [1]. In recent years, there has been no serious damage to fisheries

61

from red tide events of harmful algae, but much money is spent for the mitigation activity of red tides blooms every year. There have been many studies on red tide because the damage to sea farming by red tide blooms of harmful algae occurs every year. The research on red tide events of harmful algae in the coastal area of South Korea has mostly focused chemical properties and investigation of biological causes and is generally used for mitigation activity [2]. However, research on a variety of automatic methods for monitoring and recognizing red tide images of harmful algae on the coast of South Korea is still insufficient [1]. Recently, studies of red tide recognition mostly use classification techniques as follows. Jiang et al. proposed the red tide algae classification method using Bayesian classifier and SVM (support vector machine). Their method removes unknown and polluted algae images using Bayesian classifier. Red tide algae images are classified using SVM [3]. However, there is no good performance to remove unknown and polluted images. Besides, they use only 8 species of red tide algae for training images classification. Jiang et al. proposed red tide algae recognition method using two step classifiers. Their method uses first step classifier in which SVDD (support vector data description) removes unknown and polluted algae images. Algae images are classified into the predefined class label using second step SVM classifier. Also this method uses only 8 species of red tide algae for learning image classification [4]. Lili et al. use SVM and FCM (fuzzy c-means) clustering method for classifying red tide images. Their method uses SVM to identify class of test sample. The method identifies the misclassified data using probability based method, and then the misclassified data is reclassified using FCM. Their method also uses 5 species of red tide algae to train image classification [5]. Their methods [3, 4, 5] are difficult to classify red tide algae images except training algae images. In addition, they do not consider matching point between training images and query image, since the training images have different forms regarding various species of red tide algae (i.e., there are blooming red tides with relation to 63 species of algae in the coastal area of South Korea). In order to resolve the above limitations of few training algae images and without matching point, this paper proposes a new red tide image recognition method using roundness and entropy based on semantic features of LSA. The proposed method has following advantages. First, training of the method uses the semantic features of 63 species of red tide algae image by means of LSA which can recognize various species of red tide algae images. Second, the method uses roundness of shape of algae images to apply matching point of the images for the images recognition. Final, our method uses entropy of red tide images which improves the recognition accuracy of algae images.

2

Latent semantic analysis

We define the matrix notation as follows: Let X*j be the j’th column matrix vector X, Xi* be the i’th row vector, and Xij be the element of the i’th row and the j’th column.

62

Latent semantic analysis (LSA) is a decomposition method using singular value decomposition (SVD). This method decomposes matrix A into three matrices, U, D, and VT [6, 7]. ~~ ~ (1) A = UDV T , A ≈ UDV T where U is an m×n orthonormal matrix of eigenvectors of AAT (left singular vectors), and V is an n×n orthonormal matrix of eigenvectors of ATA (right singular vectors). D= diag (σ 1 , σ 2 ,..., σ n ) is an n×n diagonal matrix whose diagonal elements are non~

~

negative eigenvalues sorted in descending order. U is a m×r matrix, where U ij = Uij ~ ~ ~ if 1 ≤ i ≤ m and 1 ≤ j ≤ r . D is a r×r matrix, where Dij = Dij if 1 ≤ i, j ≤ r . V ~

is a n×r matrix, where Vij = Vij if 1 ≤ i ≤ n and 1 ≤ j ≤ r . Finally, r 0 (12) 𝑢, 𝑧, ℎ   ∈ 0,1 , 𝑥, 𝑤, 𝑒 ∈ 𝑅 (13) The constraints (2), (3), (11) and (12) are about activating or not the sensor nodes. The constraints (4), (5), (6), (7) and (8) are responsible for routing. The constraint (10) is responsible for sensor’s battery’s electrical charge usage limits.

3 The base hybrid methodology Both the exact and meta-heuristic approaches have pros and cons when dealing with hard combinatorial optimization problems. But their hybridization, when properly done, may allow the merging of their strong points in a complementary manner. However, the size and complexity of the optimization problems faced nowadays have increased a lot, demanding for the development of new methods and solutions that can find acceptable results within a reasonable amount of time. In this context, a hybrid methodology has been recently introduced in the literature by [3], trying to push forward the boundaries that limit the application of an exact method through the decomposition of the original problem into two conceptual levels. According to the framework underlying this approximate methodology, the exact method works no more with the original problem, but with reduced instances of it. That preserves its conceptual structure. By this means, an optimal solution to a given sub-problem will also be a feasible solution to the original problem. On the other hand, the meta-heuristic component of the framework works on a complementary optimization problem, that is, the design of reduced instances of the original problem formulated as mathematical programming models. It is referred to as the Generator of Reduced Instances (GRI), whose goal is to determine the subset of

254

points of the reducible structure that could derive the best sub-problem instance; that is, the sub-problem which, when submitted to the SRI, would bring about the feasible solution with the highest possible objective function value. In this scenario, the objective function values of the solutions that could be realized by the solver are used as figure of merit (fitness) of their associated sub-problems, thus guiding the metaheuristic search process. The interaction between GRI and SRI is iterative and repeats until a given stopping condition is satisfied. So far, the meta-heuristic chosen to implement the reduced instances generator has been a Genetic Algorithm explained by [4]. This option is due to the good levels of flexibility and adaptability exhibited by the class of evolutionary algorithms when dealing with a wide range of optimization problems as presented by [5]. The genetic representation of the individuals (chromosomes) follows a binary encoding that indicates which decision variables belonging to the reducible structure will be kept in the new sub-problem to be generated. That is, those genes having ‘1’ as alleles define the subset of variables that generates the reduced instance. Conversely, the exact method is assumed to be any state-of-the-art algorithm used to solve mixed integerlinear problems, such as Branch-and-bound or Branch-and-cut described in [6].

4 Improvements for the dynamic coverage and connectivity in wireless sensor network problem Adapting the base hybrid methodology to be suitable for a totally different class of problem is a challenge. A drawback that has limited the effectiveness of the base hybrid methodology as presented in Section 3 relates to its propensity for bringing about an uncontrolled density explosion over the individuals (i.e. reduced instances of the original problem) produced by the GRI. We define “density of an individual” as the ratio between the numbers of genes having ‘1’ as allele (referred to as activated) and its total length. The fact is that an increase in density tends to generate subproblems more closer to the original problem, thus possibly yielding better solutions. This situation can be better pictured as if having some sort of an “attractor” pushing the overall population density up as the GRI (GA) evolves. Although expected, this phenomenon may have an undesirable side effect if it occurs prematurely. This is because, usually, high densities imply higher complexity to be dealt with by the SRI, which indirectly affects the search process conducted by the GRI as the time spent in each generation tends to become progressively higher. This may cause a drastic limitation over the number of search iterations performed by the SRI, hindering both the effectiveness and efficiency of the whole optimization. 4.1 Compact chromosome encoding According to [4], the right representation of the individuals is one of the most difficult parts of designing a good evolutionary algorithm. The binary chromosome encoding was used in the original version of the hybrid methodology. Each gene represents the inclusion of the equivalent element of the

255

reducible structure that will be considered in the generation of the new sub-problem. This type of chromosome encoding is not appropriated for other problem domains like the one treated in this work. It would generate too large chromosomes. The proposed new encoding represents the integer indexes of the sensors that must be taken in the sub-problem generation. So there is no need of representing all sensors. Only a small amount of sensors has to be considered and the length of this chromosome can be down to 17% of the binary encoding one.

5 Computational Results The grid sensor placement was used for simplicity sake because the random scenario did not present significant variation of the problem complexity which is the main concern of these experiments. The machine used on this test was an Intel Core 2 Quad 64 bits with 8 GB of RAM machine with OpenSuse Linux 11.0 64 bits. As Linear Integer Programming (LIP) solver, the IlogCplex 10.1 dynamic library [7] was used attached to a Java program that implements the methodology. Table 1 presents the comparison of Hybrid Methodology (HM) and LIP approaches. On these experiments, the demand points are disposed in a grid and the last two columns present the results when demand points are spread randomly. Due to the stochastically nature of the HM, it is presented the average and standard deviation, separated by the symbol ±, of results found in a batch of 10 problem instances. The objective function is composed by the summation of electrical charge consumption in all sensor nodes and the penalties. The penalties allow uncovered demand points, giving flexibility to the model, but at the same time, avoid the unnecessary use of this resource. Thus, the real objective is calculated by subtracting the artificial coverage penalties of the objective function or just calculating the first part (summation) of the objective expression. The amount of demand points and sinks were 400 and 1 respectively. HM Time intervals 6 Sensor nodes 36 Time (minutes) 151.78 ± 14.61 Time for first 79.84 ± solution 54.65 Uncovered 0.93 ± 0.39 demand pts (%) Real objective 22,223.27 ± 2,614,63

HM 10 36

LIP 10 16

Random HM 6 36

Random HM 10 36

189.80 ± 61.01

298.55

146.77 ± 32.45

366.28 ± 13.54

104.14 ± 75.06

298.55

91.55 ± 47.14

196.05 ± 78.30

2.35 ± 0.54

0.33

1.56 ± 0.61

2.10 ± 0.53

26,665.91

20,452.51 4,135,35

33.374,04 2,389.76

±

± 32.522,51 2,628.33

±

Table 1: Simulation results for demand point in grid and randomly positions. The real purpose of this model is to extend the WSN lifetime as far as possible, preserving the WSN cost. So, lower electrical charge consumption is not necessarily an important issue if it does not reflect in more time slots. The number of time slots multiplied by the duration of each time slot represents this WSN lifetime.

256

Given this explanation is reasonable to say that both solutions found by HM and LIP are equivalent in effectiveness. However, the HM approach can handle an amount of sensors 325% times larger, extending the working range of this application. The only drawback is the uncovered demand point rate, which is worse than LIP value. Despite this imperfection of 2.35%, many real applications tolerate some lack of coverage by the nature of the observed phenomenon and other aspects.

6 Conclusion This hybrid methodology is not only suitable for solving complex instances in the domain of cutting and packing problems. It can be adapted to tackle other problem classes like WSN as shown here. The key point in this adaptation is finding the best or at least a good reducible structure. This analysis is very linked to the chromosome encoding choice as it represents a trade of between sub-problem complexity range width and chromosome size. A good reducible structure allows a wide range of sub-problem complexity from very light and fast sub-problems to the actual real problem. On the other hand the reducible matrix size affects the chromosome size and a large chromosome size reduces the GA effectiveness. In this problem a good reducible structure was found but it is much larger than the ones found in the cutting and packing problem instance. That is the reason why a new chromosome encoding was developed making the matrix choice viable. The result found is far better than reference literature and leaves opportunities of future enhancements as new supplementary algorithms and heuristics are aggregated to this methodology.

References 1. Loureiro, A., Ruiz, L., Mini, R., and Nogueira, J. (2002), Redes de sensores sem fio. Simpósio Brasileiro de Computação Jornada de Atualização de Informática. Technical Reports 030001, January, 2003. 2. Nakamura, F. G.,Quintão, F. P.,Menezes,G.C.,Mateus, G. R. (2004), Planejamento dinâmico para controle de cobertura e conectividade em redes de sensores sem fio. Workshop de Comunicação sem Fio e Computação Móvel, 182-191. 3. Nepomuceno N., Pinheiro P. R., and Coelho A. L. V.( 2006) Aplicação de uma Metodologia Híbrida ao Problema de Carregamento de Contêiner. In: XXXVIII Simpósio Brasileiro de Pesquisa Operacional, 2006, Goiânia. Anais do XXXVIII Simpósio Brasileiro de Pesquisa Operacional. Rio de Janeiro, 2006. p. 1596-1603 4. Eiben, A. E., and Smith, J. E., (2003) Introduction to Evolutionary Computing. Springer-Verlag. 5. Back, T., Fogel, D. B., and Michalewicz, Z., Handbook of Evolutionary Computation. IOP Publishing Ltd., Bristol, UK, UK, 1997. 6. Wolsey, L. A., Integer Programming. John Wiley& Sons, 1998. 7. ILOG, ILOG CPLEX 10.0 User's Manual, 2006.

257

Anatomy of Biometric Passports Dominik Malˇc´ık and Martin Drahansk´ y Faculty of Information Technology, Brno University of Technology, Boˇzetˇechova 2, 61266 Brno, Czech Republic {imalcik,drahan}@fit.vutbr.cz http://www.fit.vutbr.cz

Abstract. A recent trend in a field of personal identification documents is to use RFID (Radio Frequency Identification) technology and biometrics, especially (but not only) in passports. This paper provides an insight into the electronic passports (also called e-passport or ePassport) implementation chosen in the Czech Republic. Such a summary is needed for further studies of biometric passports implementation security and biometric passports analysis. A separate description of the Czech solution is a prerequisite for a planned analysis, because of a uniqueness of each implementation1 . Key words: passport, ePassport, electronic passport, biometric passport, RFID, chip, fingerprint, ICAO, algorithm, MRTD, Machine Readable Travel Documents

1

Introduction

The idea of a better passport system incorporating biometrics has been alive for more than 20 years. However, it has taken considerable time to prepare all aspects for the new technology. Using biometrics to improve the system of travel documents is undoubtedly a crucial milestone. Naturally, there are security threats due to the fact that all biometric features are usually very sensitive information that has to be treated appropriately.

2

Passport chip

The RFID technology was chosen for implementation of the ePassports (all the electronic passports are labelled with an international logo – see the red circle in Fig. 2). The used chips have to meet standard compliant with ISO 14443 with modulation A or B – frequency for transmissions is 13,56 MHz with a short range (max. 15 cm) [7]. The passport RFID chip provides, among others, cryptographic functions, r/w memory modules accompanied with memory modules that are readable only for the tag itself (no information from these memory cells can be retrieved out of the device). 1

Each country can choose the implementation details within a range specified by the ICAO (International Civil Aviation Organisation – http: // www. icao. int/ ); moreover, specific security mechanisms are optional and can be omitted.

258

2

2.1

Anatomy of Biometric Passports

Passport chip memory

Fig. 1. Memory data groups of passport RFID chip. Please notice especially the description of DG1-DG5 (source: [5]).

The memory is logically divided into two main regions - one is accessible from outside of the chip (via wireless communication), the second one provides a part of security by hiding its content – the hidden content is available only for internal functions of the chip. The part of memory available for reading provides sixteen separated data groups (labelled as DG1, DG2, ..., DG16 – see Fig. 1). Each group incorporates different data. Dissimilar types of protection are used over the groups of the stored data. The data groups DG1, DG2, DG3 and DG5 are important within the scope of the biometric passports, because these groups are used for storing information related to identity check. [5]

Fig. 2. On the top left: RFID chip without and with antenna; bottom left: the international logo of electronic passports; right: data page of a Czech specimen with labelled MRZ – Machine Readable Zone (source: [1]).

259

Anatomy of Biometric Passports

3

3

Use of biometrics

A proper biometric feature should be unique for each person and it should be invariable in time (usually from a specific age); given in the simplest possible way – it is an unambiguous identifier of a person. Moreover, some of the biometric features are well proven and have been even practically used for a long period of time – e.g. fingerprints in criminalistics. As it is not possible to give an exhaustive overview of biometrics, let us focus on the features that are important for contemporary passport implementation – 2D facial photo and fingerprints (the use of iris can be expected in the near future). [8] 3.1

Facial photograph

Facial photograph of an applicant is employed as a basic security element. This type of security is well known also from older types of documents. In classic paper documents, the facial photo primarily serves for visual identification by officers. Despite the officers’ training and their ability to recognise a person even if there is some change in an applicant’s appearance (moustache, haircut, glasses, etc.), the case of similar individuals (twins, siblings or even doubles) could lead to identity mismatch. If the facial photo is treated from a biometric point of view (not just as a picture of a person) – the face contains information that is invariant in time and can be measured, e.g. the distance between eyes, position of chin, position of nose, etc. These factors can affect the recognition process by providing additional information to the officer. Nonetheless, the twins will still look similar. That is why an absolutely different security component is needed (see chapter 3.2). [8] Picture data storage The picture data (facial photo) is taken according to specifications in ISO19794-5 that defines conditions for acquiring this type of data: format, scene, picture properties, etc. The picture data is stored on the chip twice (DG2 and DG5 – see Fig. 1), both in JPEG/JPEG 2000 format. The first occurrence is designated for laser engraving with following properties – grayscale, 60px distance between eyes, resolution of 620×796, stored in DG5. The second picture is encoded and stored in DG2 in full colour, resolution of 240×320 with max. size of 15 kilobytes. This smaller image is used for biometric identity check. [3], [5], [9] 3.2

Fingerprints

With respect to the facts introduced in chapter 3.1, the need for new reliable means of identity verification has been solved by introducing fingerprints. It has been proven that even fingerprints of monozygotic twins are significantly different. That means the two identities of twins can be obviously distinguished by matching the corresponding fingerprint with its stored digital representation (or

260

4

Anatomy of Biometric Passports

with a paper record of the fingerprint). Even so, there still do exist possibilities for counterfeiting fingerprints. Nevertheless, the fraudsters have to face the problems with tricking the fingerprint scanners, because the scanners are being more often equipped with sophisticated liveness detection - especially when a security risk is expected. Sometimes it is simply almost impossible to cheat the fingerprint checking, because of a presence of an officer. Adopting this measure naturally does not result in an absolutely perfect protection against unwanted actions2 . Nonetheless, the security level has rapidly increased with incorporating a fingerprint check. [8] Fingerprint data storage Fingerprints are taken in compliance with ISO/IEC FCD 19794-4 and ANSI/NIST-ITL 1-2000 IS standards. The quality of stored fingerprint has to be marked with NFIQ (NIST Fingerprint Image Quality) equal to 3 or a better grade. In Fig. 1 can be seen that a DG3 has been designed to hold fingerprint data. Maximal data size of one fingerprint is 15 kB in compressed format WSQ specified in document IAFIS-0110 (V3), precisely according to the Gray-scale Fingerprint Image Compression Specification 1997. [9], [1]

4

The Czech implementation

The Czech electronic passport was introduced as a second device of this type in the EU. Since that time new types of security has been already introduced, however due to the backward compatibility of all solutions across the world and given minimal requirements of ICAO, the former threats will be still present. Other details regarding the whole Czech project (involved organisations, communication channels, passport producers, etc.) are not relevant for this paper3 . 4.1

Security

The main goal of the whole ePassport project is to preserve privacy of the personal data and prevent forgery of the travel documents. Different measures are used with respect to the importance of the particular aspect. Mechanical and optical elements Security elements of this type are often used not only in passports (but also, e.g., bank notes, other types of personal documents, etc.). Although we do not aim at this type of security, let us mention at least some of them (not all of the listed bellow have to be necessarily employed in the last passport revision of the Czech Republic): serial numbers, fluorescent elements, relief stamping, engraving, guilloches, holograms, laser perforations, mechanic perforations, watermarks and many more. 2 3

Absolute security does not exist. This information will be published in a different paper.

261

Anatomy of Biometric Passports

5

Basic Access Control (BAC) A very simple mechanism used for protection of information stored in DG1 and DG5 (see chapter 2.1). The BAC technique is based on two crucial principles – the first: data can be read only in case the passport is opened on the data page (if not, the RFID chip is shielded – a communication cannot be technically established); the second, the MRZ contains information which is used for the transmission password derivation). Actually, the data in DG1 and DG5 are the same as the information on the passport data page (see the right part of Fig. 2), that is why in the case the attacker has the ability to open the passport on the data page and read the MRZ, the information from the data page (and also from the DG1 and DG5) is not secret anymore. The keys for the BAC are derived by SHA-14 from the MRZ, precisely from the passport serial number (9 characters), owner’s date of birth (6 characters) and the date of expiry (6 characters). The result of the hash function is truncated to 16 bytes and divided into two parts (key A: 0th-7th byte; key B: 8th-15th byte) for 3DES. A key for the main communication is then established via 3DES encoded messages. [1], [2], [3], [7] Active Authentication (AA) The active authentication serves as a protection against passport cloning. A couple of keys (private and public) is generated during the process of personalization of a new passport. The private key is stored in a part of memory that is inaccessible from outside of the chip (it is provided only in the hardware of the chip). The public key is freely available in DG15. [5] The principle is then based on the asymmetric cryptography. Random data are generated and sent to a passport chip by a reader. The data are signed internally with the private key stored in the chip and sent back to the reader. In the last step, the reader verifies compatibility of the key pair and emits a result about authenticity of the private key. [3], [1] Extended Access Control (EAC) The aforementioned BAC is definitely too weak to secure the sensitive biometric data – the fingerprints (DG3), in the future also the iris (DG4). Therefore, a new security specification was made. The EAC was specified in technical report BSI TR-03110 (Advanced Security Mechanism for Machine Readable Travel Documents – Extended Access Control). The EAC has been used in the Czech Republic since April 1, 2009 (it was brought to light together with incorporation of the fingerprints). [3], [1], [7] Two cryptographic mechanism are being used within EAC: Chip Authentication (CA, based on Diffie-Hellman) It is an alternative to Active Authentication (protection against chip cloning). In contrast to the Active Authentication, the CA does not suffer from so called challenge semantics. The challenge semantics can cause tracking of the owner’s transfer in a specific case. That is why Germany did not include AA into their implementation of ePassport. After the DH process a cryptographically strong shared secret is available for encoding the following communication. [7] 4

It is possible to use SHA-1 or SHA-2 (SHA224, SHA256, SHA384, SHA512).

262

6

Anatomy of Biometric Passports

Terminal Authentication (TA, based on PKI) Only approved terminals have permission to access the data groups with biometric data. The terminal has to be equipped with a valid certificate of particular country to access the data. Each terminal is set to a specific self destruction time period. The length of this period depends strictly on conditions of use of each terminal (from 1 shift to 1 month max.). Each terminal is labelled with unambiguous ID and can be blocked. [3], [1]

5

Conclusion

This summary of the ePassport implementation in the Czech Republic will be used as a basis for the next steps in an analysis of hardware (microscopic analysis, side channel analysis, etc.) and software (protocols analysis, firmware analysis, etc.) of such passports that will be performed within the next months.

6

Acknowledgment

This work is partially supported by the research plan ”Security-Oriented Research in Information Technology”, MSM0021630528 (CZ), by the grant ”Advanced secured, reliable and adaptive IT ”, FIT-S-11-1 (CZ), by the grant ”Information Technology in Biomedical Engineering”, GD102/09/H083 and by the project ”The IT4Innovations Centre of Excellence”, ED1.1.00/02.0070.

References 1. Holenda, T.: Prezentace projektu ePas pro odbornou konferenci Systemov´ a Integrace (Presentation of the project ePassport for conference System integration), presentation, Ministry of The Interior of the Czech Republic, 2009. 2. Holenda, T.: Odborn´ a konference Quality & Security (Conference Quality & Security), presentation, Ministry of The Interior of the Czech Republic, 2007. ˇ 3. Mayer, P.: Biometrick´e pasy v Cesk´ e republice (Biometric passports in the Czech ˇ 2007. Republic), presentation, Siemens CR, 4. European Commission: Borders & Visas - Document security, Jan 2012. http:// ec.europa.eu/home-affairs/doc\ centre/borders/borders\ doc\ en.htm 5. The International Civil Aviation Organisation: Machine Readable Travel Documents (Part 1, Volume 2), ICAO, 2006, ISBN 92-9194-757-1. 6. Sheetal, S.: Technical analysis of security mechanisms used in RFID E-passport, related threats, security and privacy issues, Viterbi School of Engineering, University of Southern California. 7. Raˇsek, L.: Elektronick´e pasy – jak funguj´ı (Electronic passports – how it works), Viterbi School of Engineering, Czech Open System Users’ Group, Conference proceedings, 2006. 8. Drahansk´ y, M., Ors´ ag, F., Doleˇzel, M. et al.: Biometrie (Biometrics), Brno, CZ, Computer Press, 2011, s. 294, ISBN 978-80-254-8979-6. 9. Maleˇc, F.: Druh´ a generace elektronick´ ych pas˚ u a nov´ a generace elektronick´ ych pr˚ ukaz˚ u o povolen´ı k pobytu (The second generation of electronic passports and a new generation of electronic documents), presentation, SmartCard Forum, 2010.

263

A Novel Adaptive Virtual Machine Deployment Algorithm for Cloud Computing Hongjae Kim1, Munyoung Kang1, Sanggil Kang2, Sangyoon Oh1 Department of Computer Engineering, Ajou University, Suwon, South Korea Department of Computer Science and Information Engineering, Inha University,Incheon, South Korea 1 {carrotbox, hanamy, syoh}@ajou.ac.kr 2 [email protected]

Abstract. Virtualization is one of the key enabling technologies for Cloud Computing. When we utilize this technology to abstract physical resources such as memory and CPU for flexible use of them, a virtual machine deployment algorithm does essential role for improving efficiency and load-balancing. We propose a new adaptive VM deployment algorithm based on the Hungarian algorithm to support a concurrent deployment with multiple virtual machine instances. By using the algorithm along with a secondary job queue in the enabling architecture, the load balancing performance can be improved as well as the overall utilization. Experimental results of ours show the significant improvement in load balancing over the general Multi-Dimensional Bin Packing algorithms. Keywords: Virtual Machine, Job Deployment, Bin Packing Problem, Hungarian Algorithm

1

Introduction

As Cloud Computing becomes a de facto standard for computing, Infrastructure as a Service (IaaS) has been emerged as an important paradigm in IT area. By applying this paradigm, we can abstract the underlying physical resource such as CPUs, memories and storages and offer these virtualized resources to users in the form of Virtual Machine (VM). Multiple VMs are able to run on a single Physical Machine (PM). Because of these benefits: low cost, flexibility, and manageability, IaaS becomes more popular for a data center solution, a resource renting, and a method for legacy system integration. A typical use of IaaS is as follows; a user requests VMs for their use. Then a provider such as Amazon, the largest IaaS provider, creates VMs on PMs and provides the created VMs to users. The VM deployment process maps VMs to PMs. In this process, a PM’s computing resource (i.e. capability) and a required VM’s resource (a demand) are two major inputs of the placement problem. A wrong VM deployment makes inefficient use of computing resources (i.e. VMs are mapped to PMs which are not suitable) which causes low resource utilization and imbalanced job loading. Thus, the performance of the VM deployment process is important factor of

264

efficiency and load-balancing performance of Cloud service. Many researchers have been studied about the VM deployment problem. To model the problem, the Bin Packing has been used in many studies [1, 2]. In Bin Packing problem, PMs are regarded as bins which have the capacity and VMs are regarded as items which have weight and placed in the bins. However, in many cases of Bin Packing problem modeling, only one VM request is mapped to PM at a time that leads to its limited packing and load balancing capability. To address this ‘one-by-one’ deployment problem, we propose a new adaptive VM deployment algorithm. In this algorithm, we define the score which represent the utilization of each PM. Then we extend the Hungarian algorithm [3] that solves the assignment problem. With it, we can find an optimal assignment solution of n VM requests to n PMs based on the calculated score. In our proposed algorithm, multiple VM requests can be mapped to PMs simultaneously to improve the load balancing performance of given system. We introduce a secondary job queue in the enabling architecture for this simultaneous deployment. In the experiments, we evaluate the standard deviation of the PMs. The standard deviation is used in many researches to show the load balancing performances [4, 5]. Our experiment results show the apparent performance improvement compared to the general Multi-Dimensional Bin Packing algorithms. The remainder of this paper is organized as follows; we describe related works of this research in Section 2 and describe our adaptive VM deployment algorithm in Section 3. We present the evaluation of the proposed algorithm in Section 4 and we conclude in Section 5

2

Related Works

The Bin Packing is the problem of finding the assignments of items with the weight to bins with the capacity. We may draw an analogue between the item-bin and the VMPM. Also the weight can be regarded as a required VM resources and the capacity can be regarded as a total resources of PM. Thus the Bin Packing problem may be a good tool to model the VM deployment problem [1, 2]. However a PM has multiple computing resources and the conventional Bin Packing problem cannot model multiple resources. Thus multi-dimensional Bin Packing Problem should be used for those cases. There are two well-known algorithms for the Bin Packing Problem: First Fit Decreasing (FFD) and Worst Fit Decreasing (WFD). In the FFD, requested VMs are sorted by required resources in decreasing order. Then, each VM is orderly mapped to the first PM which has enough resources. On the other hand, VMs are sorted by resources in decreasing order in the WFD. Then, each VM is orderly mapped to the PM which has most free resources. Therefore FFD can deploy more VMs than WFD and WFD can distribute loads more efficiently than FFD. However, both algorithms can map only one VM request to a PM at a time. Since multiple VMs can be mapped to one PM, it will be resulted in a low resources utilization and imbalanced load. We extend the Hungarian algorithm for a VM deployment. The Hungarian algorithm for the assignment problem was proposed by Kuhn [3]. Using it, we can find the optimal assignment of n jobs and n machines. The optimal assignment is the

265

one makes the sum of the setup time for their assigned machines as a minimum. We can think of a job as a VM and machine as a PM. Likewise, a setup time can be replaced with a utilization of PMs. The Hungarian algorithm solve the problem in O(n3) time and FFD solve the problem in O(nlogn) time. However, the Hungarian algorithm can assign multiple VMs to PMs simultaneously and it can distribute loads more efficiently than FFD and WFD.

3

Adaptive VM Deployment Algorithm

In the enabling architecture of our proposed algorithm, the front-end node plays a mediator role between users and physical resources (as shown in the Fig. 1). It receives VM requests from users and distributes the requested VMs to PMs based on the proposed algorithm. There are two software components in the mediator: a monitor and a scheduler. The monitor component is responsible for measuring utilization of each PM. The scheduler is responsible for receiving VM requests and distributing them to appropriate PMs.

Fig. 1. Architecture overview of the IaaS front-end node.

In the enabling architecture, we introduce a secondary job queue in addition to a primary job queue for incoming job requests. The proposed algorithm is applied for distribution of job in the primary queue. While the job in the primary queue is in distribution, the secondary queue is holding incoming VM requests from users. The size of secondary queue is static and the one for the primary queue is dynamic. Since the size of primary queue should be less than the number of PMs that have enough resources for VM requests, the primary queue cannot store the larger number of coming VM requests than the number of PMs. That leads to a necessity of having a secondary queue to keep the incoming VM requests. The front-end node sends the incoming VM requests in the secondary queue. VM requests are transferred to the primary queue if it is empty. Then, the front-end node distributes VM requests stored in the primary queue to PMs by using the proposed adaptive VM deployment algorithm. In the proposed VM deployment algorithm, we consider the VM execution time and two computing resources, a CPU and a memory to calculate the score. A PM i is represented by pi(pci, pmi), where pci is the CPU capacity and pmi is the memory

266

capacity. A VM j is represented by vj(vcj, vmj, vtj), where vcj is the required CPU, vmj is the required memory of VM j and vtj is the execution time of the VM. Si represents the VMs of pi. If there are m VMs in a pi, Si represented by Si = {v1, v2, v3, …, vm}. We define the score parameter at time t and we use this parameter to project resource usage of pi in its execution time. The proposed algorithm uses the score to distribute VMs to PMs for balancing the resource utilization. s

( )=∑

×

∈

×

−

(1)

Now we can get the standard deviation of the overall score among PMs. Let scorestdev denotes the calculated standard deviation of score. To choose the target PM where the VM request will be deployed, we make scorestdev matrix where m VM requests (rm) with i th PMs (pi) as shown in the Table 1. In the matrix, scorestdev mi denotes the standard deviation of PMs when VM request rm deployed to PM pi. After all these calculations, we deploy VM request to PM which have minimum scorestdev value and it will make the load balanced in the resource utilization. Table 1. The scorestdev matrix where m VM requests with i PMs.

r1 r2 r3

p1 scorestdev 11 scorestdev 21 scorestdev 31

p2 scorestdev 12 scorestdev 22 scorestdev 32

p3 scorestdev 13 scorestdev 21 scorestdev 33

ㆍㆍㆍ

ㆍㆍㆍ

ㆍㆍㆍ

ㆍㆍㆍ

rm

scorestdev m1

scorestdev m2

scorestdev m3

ㆍㆍㆍ ㆍㆍㆍ ㆍㆍㆍ ㆍㆍㆍ ㆍㆍㆍ ㆍㆍㆍ

pi scorestdev 1i scorestdev 2i scorestdev 3i scorestdev mi

Our proposed algorithm, like Hungarian algorithm, consists of 4 steps; Step 1: Find the minimum scorestdev value in each row and subtract off. Step 2: Find the minimum scorestdev value in each column and subtract off. Step 3: Draw as few line as possible to cover all the zeros in the matrix. If the number of lines is less than the number of VM requests, find a uncovered minimum scorestdev value. Then, subtract the uncovered minimum scorestdev value from every uncovered scorestdev value and add the uncovered minimum scorestdev to every scorestdev values that are covered with two lines. Then, repeat step 3. If the number of lines is equal to the number of VM request, go step 4. Step 4: From the top row, make an assignment. The assignment is made when there is only one zero in a row.

4

Evaluation of the Algorithm

In order to prepare the VM and PM type, we refer the Amazon EC2 Instance Types [6] and obtain ten types of VM and one type of PM. VMs have different CPUs and memory requirements depending on the type. The given types are (1, 2), (2, 4), (4, 8), (6, 18), (8, 2), (8, 16), (10, 4), (12, 4), (18, 6), (20, 6). For example, type (1, 2) mean that it require

267

1 CPU and 2GB memory. There are four kinds of execution time for each VM; 4, 8, 12, and 24 hours. During experiments, each VM’s type and execution time are determined randomly. Also, we consider 50 PMs with 80 CPUs and 60GB memory. We set up for two test cases. The test case 1 is to evaluate load balancing performance with small VM requests per hour; between 10 and 20 VMs. The test case 2 is to evaluate load balancing performance with variable VM requests per hour; between 10 and 40 VMs.

Fig. 2. Standard deviation results with test case 1.

Fig. 2 depicts the experimental result of load balancing performance of our proposed algorithm. It shows the standard deviation of CPU utilization (a) and memory utilization (b) with test case 1. All VM requests are deployed to PMs and PMs have enough resources for deploying new VM requests when VM requests are generated. We can see from this result that the proposed our algorithm distributes given loads more efficiently than FFD or WFD. However, there is no congestion in queues.

Fig. 3. Standard deviation results with test case 2.

Fig. 3 shows standard deviation results with test case 2. We can observe from these results that standard deviations of proposed algorithm are lower than both FFD and WFD. Standard deviation is increasing at the beginning, around 40, and 100 (hour).

268

This is because deployed VMs are relatively small compared to PMs. Standard deviation is decreasing around 10, 70, and 130 (hour). This is because VMs are starting to deploy to PMs and free resources of PMs are reduced. Queue congestions (i.e. there are waiting jobs in the queue) are occurred at 20, and 80 (hour). At that time, PMs did not have enough resources for deploying VMs in queue. So, standard deviation is lower than other time.

5

Conclusion

A wrong VM deployment algorithm puts VMs to not-suitable PMs or deploys many VMs to the limited PMs. This may be resulted in a low resources utilization and load imbalance. We propose a new VM deployment algorithm which is designed to provide a simultaneous multiple VM deployment. By extending the Hungarian algorithm for simultaneous deployment as well as introducing a secondary job queue to the architecture, we can achieve better performance and stability of given system (i.e. prevent hotspot and distribute loads more efficiently). This algorithm and enabling architecture is particularly important for a cloud computing management, but there is great potential for other environment with a job deployment. To evaluate the performance of the proposed algorithm, we conducted experiments with two test cases. Our experimental results show the apparent performance improvement and more balanced job load compared with the general multi-dimensional Bin Packing algorithms. We expect the same performance increase if we apply the proposed algorithm to any cloud computing tool kits. Acknowledgement. This work was jointly supported by the MKE, Korea under the ITRC support program supervised by NIPA (NIPA-2012-(H0301-12-2003)) and Basic Science Research Program through the NRF of Korea (No. 2011-0015089).

References 1. Wilcox, D., McNabb, A., Seppi, K., Flanagan, K.: Probabilistic virtual machine assignment. In: CLOUD COMPUTING2010, 1th International Conference on Cloud Computing, GRIDs, and Virtualization, pp. 54--60. IARIA, Lisbon (2010) 2. Hyser, C., Mckee, B., Gardner, R., Watson, B. J.: Autonomic virtual machine placement in the data center. Technical report HPL-2007-189, HP Laboratories (2008) 3. Kuhn, H. W.: The Hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2, 83--97 (1955) 4. Ni, J., Huang, Y., Luan, Z., Zhang, J., Qian, D.: Virtual Machine Mapping Policy Based on Load Balancing in Private Cloud Environment. In: International Conference on Cloud and Service Computing (CSC). pp. 292--295. Hong Kong (2011) 5. Zhou, S.: A Trace-Driven Simulation Study of Dynamic Load Balancing. J. IEEE Transaction on Software Engineering, 14, 1327--1341 (1988) 6. Amazon EC2 Instance Types, http://aws.amazon.com/ec2/instance-types

269

The Design of a Web Contents Management System Jaegeol Yim, Gyeyoung Lee Computer Engineering, Dongguk University at Gyeongju {yim, lky}@dongguk.ac.kr

Abstract. According to analysis at SNL Kagan, the market share of IPTV is rapidly growing whereas the market share of cable TV is slowly shrinking. This paper introduces an ‘Open IPTV System’ which allows a subscriber to easily open his or her own IPTV broadcasting station. In the ‘Open IPTV System’, operators must be present at the office to access the content management system (CMS). We implement a web CMS so that the operators can access the CMS anytime and anywhere as long as they are connected to the Internet. Keywords: IPTV, Contents Management, Web System.

1

Introduction

The IPTV market share is rapidly growing whereas the cable TV market is shrinking. According to an analysis done by SNL Kagan, the number of IPTV subscribers was 40 million, which is 6.1% of all pay broadcast subscribers in the world, at the end of 2010; the IPTV revenue in 2010 was $12.9 billion US, 142% of the IPTV revenue from 2009. In the light of the prosperity of the IPTV business, we have created the ‘Open IPTV System’. It is said to be open because it allows subscribers to upload their own content and easily open their own IPTV broadcasting stations. One of the disadvantages of the system is that operators must be sitting in a chair in front of the desktop computer in the computer room in order to run the contents management system (CMS). The main purpose of this paper is to implement a web CMS so that an operator can run the CMS anytime and anywhere as long as he is connected to the Internet. The ‘Open IPTV System’ is described in Fig. 1. The main components of the system are the Ingest system, CMS, SV-Solution, and the ‘DB and NAS’ system. The Ingest system is used to encode and transcode the source contents into the IPTV format. For example, an analog video tape needs to be encoded into a digital codex before being transmitted through the Internet. The main roles of the CMS are creating digital multimedia contents and saving and retrieving the contents. The SV-Solution is our broadcasting system. If a subscriber requests a video service from a TV, a smart phone, or a PC, then the SV-Solution streams out the video to the subscriber. The SVSolution consists of many server computers, such as streaming servers, IPTV middleware servers, mobile distributors, and so on. The ‘DB and NAS’ represents the storage system of the system. Multimedia files are saved in the NAS and the metadata

270

for the files are saved in the DB. For this system, this paper implements a web system that allows operators to manage the contents from remote locations.

Fig. 1. The main components of the 'Open IPTV System'

2

Related Works

This paper develops a web CMS for the ‘Open IPTV System’. Therefore, the IPTV system, contents management, and the web system are included in the related topics. In [1], the authors illustrated how IPTV would be adopted to wireless and mobile networks. Technical obstacles, such as capability limitations, bandwidth limitations, vulnerable wireless links, coverage implications, and so on were discussed. The existing IPTV architecture dedicates a centralized regional station to serve the subscribers in their respective region. This causes a significant imbalance in resource utilization. In [2], the authors proposed a novel request dispatching protocol that allows IPTV stations of different regions to collaboratively serve user requests. Quality of Service is one of the key factors to the success of the IPTV businesses. In [3], the authors depicted an end-to-end IPTV network management architecture and developed a unified network management system that monitors and controls the multi-vendor multi-technology broadband network elements. The in-house network management system supports a loosely-coupled systems integration and automatic collection of IPTV-related data from set-top boxes. IPTV service is widely believed to be the next killer application on the Internet. It is obvious that the success of an IPTV business will depend on the ability of the service providers to provide the right IPTV content and services to the right subscribers at the right time. The authors of [4] propose a business model in which service providers collect content and services and package them along with advertisements to generate the correct IPTV content and services to the right subscribers.

271

If User Created Content (UCC) is allowed to be a part of IPTV, then the IPTV format is said to be open. One of the challenges in realizing open IPTV is rights protection. To address this challenge, the authors of [5] designed a packaging content authoring and consumption system for IPTV application services based on open IPTV. One of the major features of web 2.0 is in its support of more interactive capabilities than web 1.0. The authors in [6] address the relationship between web 2.0 and IPTV service; an IPTV electronic program Guide (EPG) should use Really Simple Syndication (RSS) to support personalized service. Web 2.0 tagging should be utilized to efficiently categorize IPTV content, an IPTV system should use Web 2.0 technology to easily monitor user-generated content (UGC) to avoid offensive content and copyright infringement and only post UGC that is relevant to the IPTV service. The IPTV EPG should use AJAX to make it more responsive; mash-up and open APIs should be used in IPTV systems.

3

The functional user requirements for the CMS

The CMS functions are categorized into SYSTEM, SOURCE, CONTENT, MONITORING, WATCHLOG, and STATISTICS, as shown in Fig. 2. Each of these categories is described further in this section.

Fig. 2. The CMS main menu

‘Management of users’, ‘management of authority’, and ‘management of code’ are submenus of SYSTEM. The ‘Management of users’ provides the functions of displaying user (subscribers) information, searching information with the user ID, creating and registering a new user, updating user information, and deleting a user. User information includes the user ID, name, description, e-mail address, office phone number, mobile phone number, and Detail. Detail is a pop-up button and a web page containing detailed information about the user is linked to it. Changing password and

272

authority level can be done on this page. There are many sorts of authorities: basic authority, administrator’s authority, authority to register and delete assets, authority to edit metadata, and delivery authority. The ‘Management of authority’ lists many different authorities so that an administrator can make check marks on the list to grant the authority to the designated user group (not to an individual user). The last subcategory of SYSTEM is the ‘management of code’. This subcategory provides the function of specifying whether or not each of the content items is available. SOURCE has ‘management of suppliers’ and ‘management of tapes’ as its submenus. ‘Management of suppliers/tapes’ provides functions used to manipulate the information of the subscribed organizations/tapes. A subscribed organization usually opens its own IPTV broadcasting station using the Open IPTV System and uploads its own content to be made available to the end users. The functions provided by ‘management of suppliers’ and ‘management of tapes’ are both similar to those provided by ‘management of users’. They are used to display, search, retrieve, update, and delete information. CONTENT has only one submenu, ‘management of contents’. The functions provided by this submenu are also used for displaying, searching, retrieving, updating, and deleting content information. Content information includes title, length (running time), status (available or not), details, preview, and delivery. The details item is similar to the details item used in the user information menu, however this one regards the content format and thumbnail items. Preview plays a low capacity version of the content when it is clicked. Delivery transcodes the content into the H.264 format (for IPTV delivery) and saves it in the BMS (Broadcast Management System) content storage. ‘Monitoring transcoding’ and ‘monitoring delivery’ are submenus of MONITORING. They manipulate the information in the transcoding and delivery functions, respectively. Delivery information includes the registration status (done or not), error type, and status of delivery. We can request redelivery here. As the others do, these submenus also display, search, retrieve, update, and delete information. WATCHLOG has ‘Tape rent’ and ‘history of delivery jobs’ submenus. The information of delivery jobs include the title of the content, the name of the file stored in the BMS storage, the name of the operator, and the date. STATISTICS includes the numbers of subscribers, subscribed organizations and content items, graphs representing the number of content items for each of the categories, information of content items open to end users, and information of content items closed to end users. The DATA option is available on this menu. It shows the statistical information regarding the designated day, week and month.

4

The Design

As was mentioned in Section 1, the ‘Open IPTV System’ described in Fig. 1 shows that the CMS is already installed and running. The purpose of this paper is to design and implement a web CMS so that operators can access it through the Internet. That means that we should use the database system and library functions in the CMS in our implementation. Therefore, we designed the structure of our system to have four

273

layers, as shown in Fig. 3: the Data Access Layer, Business Layer, Service Layer, and Presentation Layer, so that we can use the CMS database as it is.

Fig. 3. Our system design strategy

Fig. 4. The system components

We used Visual Studio 2010 as the integrated development environment (IDE) for the implementation. Windows 2008 was installed on the server and Internet Information Server (IIS) 7.5 was run on it. A SQL Server 2008 was used as the database management system (DBMS), Microsoft .NET Framework 4.0 was used for the windows programming, ASP .NET was used for the web programming, and C# was used for the general programming. In the CMS shown in Fig. 1, Syntek .NET, which is a library of database access functions, is already included, and so we used it in our implementation, as shown in Fig. 4. For the other database accesses that are not supported by Syntek .NET, we used ADO .NET. Microsoft provides the Windows Communication Foundation (WCF) as a service oriented application software development framework. We use WCF for the communication between the server and the client.

274

Recalling the user requirements, we notice that displaying a list of items is done in many places. Listing subscribers, listing contents, listing sorts of authority, listing providers are just a portion of the jobs requiring item lists. Therefore, we made a template user interface for listing items. These are the components of our system; they are listed in Fig. 4. Again recalling the user requirements,, we conclude that the functions provided by the menus and submenus can be summarized as a CRUD (create, read, update, and delete) system.

5

Conclusion

IPTV is well known to be the next killer application on the Internet. CMS is one of the most important components of an IPTV system. A web CMS is necessary in order to allow operators access the CMS anytime and anywhere. The functional user requirements analysis and design of a web CMS system has been discussed. Our design is 4 layered so that we can use the existing database system as it is. We use WCF so that the web services we implement are flexible, i.e., they can be invoked by any devices. Acknowledgments. Yim's research was supported by the ‘Industry, Academy and Research Institute Co-development of Techniques’ funded by the Small and Medium Business Administration (R00046281). Lee's research was supported by ‘Development of Global Culture and Tourism IPTV Broadcasting Station’ Project through the Industrial Infrastructure Program for Fundamental Technologies funded by the Ministry of Knowledge Economy(10037393).

References 1. Park, S., Jeong, S., Hwang, C.: Mobile IPTV Expanding the Value of IPTV. In: Seventh International Conference on Networking, pp.296-301, (2008) 2. Meng, S., Liu, L., Yin, J.: Mobile Scalable and Reliable IPTV Service Through Collaborative Request Dispatching. In, IEEE International Conference on Web Services, pp. 179-186, (2010) 3. Hu, C., Hsu, Y., Hong, C., Hsu, S., Lin, Y., Hsu, C., Fang, T.: Home network management for IPTV service operations - A service provider perspective. In: Proceedings Network Operations and Management Symposium Workshops, pp. 1-7, (2010) 4. Obele, B., Han, S., Choi, J., Kang, M.: On building a successful IPTV business model based on personalized IPTV content & services. In: 9th International Symposium on Communications and Information Technology, 809-813, (2009) 5. Kim, K., Lee, Y., Yoo, J., Lyu, W., Jung, H.: The design of the packaging contents authoring and consuming system for IPTV application service. In: The 12th International Conference on Advanced Communication Technology, pp. 586-590 (2010) 6. Kim, S., Lee, S.: Web Technology and Standardization for Web 2.0 based IPTV Service. In: 10th International Conference on Advanced Communication Technology Vol. 3, pp.17511754, (2008)

275

The Functional User Requirement Analysis of a Web Broadcasting Management System Gyeyoung Lee, Jaegeol Yim Dept. of Computer Engineering, Dongguk University at Gyeongju Korea {lky, yim}@dongguk.ac.kr

Abstract. One of the most successful examples of convergence is Internet Protocol Television (IPTV), a convergence of communication and broadcast methodologies. In this study we introduce the ‘Open IPTV System’ that allows subscribers to easily open their own IPTV broadcasting station. We then investigated the broadcast management system(BMS) of the ‘Open IPTV System’. Finally, we analyze the user requirements of a broadcast management web system that allows the operators to access the BMS anytime and anywhere as long as they are connected to the Internet. Keywords: IPTV, Broadcast Management, Web System.

1

Introduction

Convergences, such as system convergence, technological convergence, bioconvergence, and IT-convergence are some of the most popular buzzwords in the academic, research, and industrial fields. One of the most successful examples of convergence is the Internet Protocol Television (IPTV) system, which is a convergence of communication and broadcast methods. IPTV is defined by the ITU-T FG IPTV as: “IPTV is defined as multimedia services such as television/video/audio/text/graphics/data delivered over IP based networks managed to provide the required level of quality of service and experience, security, interactivity and reliability” [1]. Based on this definition, we can conclude that the smart TV is also a kind of IPTV. If someone coins a new term such as ‘intelligent TV’ or ‘active TV’ to indicate an innovatively improved version of smart TV it will also be a kind of IPTV as long as it is a multimedia service delivered over IP based networks. An IPTV system usually provides live television, time-shifted programming, and video on demand (VOD) [2]. The live television service captures the inbound TV broadcast signals in real time and encodes it into digital format for IP transmission. The time-shifted programming service allows operators to retrieve digital content, to make playback schedules, and to assign the schedule onto a channel. The VOD service allows users to select and watch a video. These IPTV services are so attractive that IPTV is starting to dominate the TV market.

276

An IPTV system consists of many component systems, including linear encoders, linear broadcasts, content management servers, data servers, nonlinear encoders, nonlinear broadcasts, and client portal servers. This implies that delivering an IPTV system demands a huge budget. The ‘Open IPTV System’ shown in Fig. 1 is an IPTV system that allows individual or collective subscribers to easily open their own IPTV stations using its infrastructure.

Fig. 1. The 'Open IPTV Stsrem'

In Fig. 1, ① represents the IP stream from other IPTV systems. The ‘Open IPTV system’ captures the IP streams from other IPTV systems, transcodes them into the H.264 format if needed, and then transmits them to a router that is the connection point to the outside world. ① represents the line from the encoder to the router. ① represents the multimedia contents(video and audio source) such as HD/SD SDI, HDMI, CVBS, SPDIF AES/EBU digital cameras, DVD, VTR, media players, the PC terminal, satellite, cable IP, DTV-STB, and video switchers. A video or audio source is either encoded into H.264 in real-time by a real-time encoder or edited by the operators using non-linear editors for delivery. ① represents that the outputs of step ① are transmitted to the broadcast management system(BMS). The end users can access the BMS and enjoy the IPTV services. ① represents the N-screen open transmission platform(mobile, PC, STB). ① represents that the administrators monitor and manage the IPTV system by accessing the open transmission platform and the router. The BMS determines whether or not the users have permission to watch the designated content, to open their own IPTV station, and to manage the contents. If the permission check is positive, then the BMS allows the user to enjoy the content, and updates the bill for the user. If the user has permission to open his own IPTV station, then the BMS allows the user to register new content, retrieve and update his content information, and delete existing content. Similarly, if the user has permission to manage the content, then the BMS allows the user to register, update and delete the IPTV content. Considering the above mentioned roles of the BMS, we can conclude that the BMS is very important and must run continuously. As shown in Fig.1, the administrators have to be present in the control room in order to run the BMS. This study has

277

developed a broadcast management web system so that the administrators can monitor and manage the ‘Open IPTV System’ anywhere and anytime, as long as they are connected to the Internet.

2

Related Works

This paper develops a web BMS for the ‘Open IPTV System’. Therefore, the IPTV system, broadcast management, and web system are included in the related topics. Among the IPTV services, namely live television, time-shifted programming, and video on demand (VOD), VOD is the most popular. The authors of [3] proposed the ‘collaborative delivery of content’ strategy. The basic concept of the strategy is to have many regional mirror stations regarding the IPTV contents instead of only keeping the content at the central station. When a user’s request arrives at the regional station, this station may ask other stations for help if it is too busy. The authors of [4] proposed a new strategy regarding home network management for IPTV service. Their main concern was in guaranteeing quality of service. For this purpose, their strategy gathers a variety of IPTV-related data from the set-top boxes and then to provide some useful management functions (e.g., video/audio packet loss analysis and IPTV audience rating) to ease service assurance and the exploration of subscribers’ tastes for IPTV programs. The authors of [5] proposed a new definition for the real-time transport protocol (RTP) system target decoder (STD) that can be used to converge the broadcast and the Internet world in IPTV. The advantage of the new definition is that the difference between MPEG2-TS and the new definition is minimal. That is, the broadcast signals can be converted into IPTV signals with a minimum of effort. Before the multimedia broadcast service (MBS) can transmit its service flow in a mobile WiMAX system, the connection ID is required. Therefore, mobile IPTV users always need to perform two processes before they are able to view a program channel: perform the IGMP join/leave at network layer and obtain the connection ID at the MAC layer. The authors of [6] proposed a new extended IGMP protocol which can be used in the mobile WiMAX radio access network in order to let the mobile IPTV services reduce their channel change time on the network. Cost-efficiency, operation flexibility and high bandwidth are attractive features that the next generation network has to provide to service providers. In order to achieve these characteristics, the network needs to have Quality of Service (QoS) provisioning abilities. As a means of providing QoS, the authors of [7] introduced a topology-based hierarchical scheduler scheme, which controls the incoming traffic at the edge of the network based on the network topology.

3

The Functional User requirements for the BMS

The functions of BMS are enumerated in Fig. 2 and are categorized into two groups: ‘basic functions’ and ‘Broadcast Functions’. The basic functions include ‘Register Content’, ‘Modify the Content’, ‘Service Statistics’, and ‘Management Linked

278

Contents’. Broadcast functions include Category, MDF(Group), URL/IMG/Live/VoD, Subscribers, Content Price, Notices, News, Administration Manager, Management log, and Setup for Client Service.

Fig. 2. The menu items

‘Register content’ menu provides functions to display content information, searching information with content description, creating and registering new content, updating content information, and deleting content. We can register a content file as it is originally or after encoding it into the h.264 protocol. When a content file is registered, the group and the server to which it belongs can be designate and a metadata aspect can also filled out after uploading the content file. The content information includes content ID, content name, description, and the manager of the content. Updating content information changes the existing content information. Deleting content deletes a content file and the metadata associated with it. The ‘Modify the content’ function allows administrators to register, update, and delete metadata. The metadata includes content title, genre, length, copyright, the thumbnail, producer, and actors. It opens a content search window when it is clicked. In this window, we can select the metadata particulars and update or delete them. Deleting metadata does not delete the associated content file. ‘Service Statistics’ function provides statistical information in regards to the following. - Contents: For each broadcast type, each content category (Live, VoD, Image, and URL), and each of the operators, this menu provides the number of content items belonging to it. - Stream server: For each of the weekdays and for each of the stream servers, this menu provides the number of visits and the number of videos being watched. - Subscribers: For each of the age types and gender types, this menu provides the number of visits and the number of watched videos. - Content provider: For each of the content providers, this menu provides the number of active (or inactive) content items provided by the content provider. - Group: For each of the groups, this menu provides the number of content items and

279

number of subscribers. The ‘Management Linked contents’ menu allows the administrators to register and update the related content items. After watching a video, a subscriber can conveniently continue watching videos related to the original. The ‘Advanced settings’ menu provides functions to design a menu (user interface) regarding the main home page of the IPTV station opened by a subscribed user of a subscribed group. For example, if a city (for example Chicago) wants to open its own IPTV station, it can open a ‘Chicago IPTV station’ under the ‘Open IPTV System’. The ‘Category’ menu allows the Chicago implimentors to easily construct their home pages. The structure of the main home page of the implemented station (the Chicago IPTV station) of the subscribed group is two-layered: the main menu and the submenu. Using the Category function, the subscribed organization (Chicago) can specify the main menu items and submenu items. For the submenu, they (Chicago, for example) can register posters that appear on the end-user’s interface. The ‘MDF(group)’ menu allows the operators to register, update and delete a group. For each of the groups we can designate a stream server, register a poster to be used as the background image, and register the RSSs(really simple syndication). The ‘URL/IMG/Live/VoD Server’ menu allows the operators to register, update, and delete a server to each of the URL/IMG/Live/VoDs. Using this menu, we can also designate a server for each of the groups. If a VoD/Image streaming server is assigned to a group, a content item is uploaded into this server when the content item is registered. The ‘Subscriber’ menu provides the functions of registration, updates and the deletion of subscribers. When it registers a new subscriber, it records the ID, password, birth date, and gender. This recorded information is utilized by the statistical, login system and billing system. Referring to this information, the login system allows or disallows the access to the IPTV system and establishes the length of time allowed to the user to access the system. This menu also takes care of the prepayment system. The ‘Price(Billing)’ menu allows the operators to register, update, and delete the content price table. For each of the content items the price can be specified and the length of the time period during which watching is free. The ‘Notice’ menu allows the operators to register, update, and delete a notice. A notice can be public to all subscribers or to a specific group of subscribers. This menu can selectively display notices based on the title or the content. This menu also displays the posters. The ‘News’ menu allows the operators to register, update, and delete news. Similarly to the notice menu, a news item can be public to all subscribers or to a specific group of subscribers. Unlike a notice, the news is delivered to the users in a real-time manner. This menu also displays posters together with titles and content. The ‘Administration Manager’ menu allows the operators to register, update, and delete administrators with permissions (content registration, category registration, management of subscribers, and so on) allowed to the administrator. ‘The ‘Management log’ menu records all of the jobs performed by the administrators of the system. Referring to this record, a manager can determine who is overloaded and whose content is registered on a given date.

280

The ‘Setup for Client Service’ menu provides the following function: for each of the news/notice items it can be specified whether or not this item should be published to the subscribers. When a new news/notice is created the reload time can be specified. The reload time is the moment when the news/notice item is published. We can also determine the subscribers to whom this news/notice is open. Using this menu, we can quickly inform the subscribers of newly available content.

4

Conclusion

The BMS of an IPTV system is taking charge of handling user access, delivering contents to the user's terminal, guaranteeing quality of service and billing. The BMS should run all the time 24 hours 7 days. In order to run the BMS, operators must be present at the control room of the IPTV center. This paper introduced broadcast management web system (BMWS) with which an operator can run the BMS anytime and anywhere as long as the operator is connected to the Internet. The detailed functional user requirements of the BMWS are discussed. In the near future, the design and implementation of BMWS will be introduced. Acknowledgments. Lee's research was supported by ‘Development of Global Culture and Tourism IPTV Broadcasting Station’ Project through the Industrial Infrastructure Program for Fundamental Technologies funded by the Ministry of Knowledge Economy(10037393). Yim's research was supported by the ‘Industry, Academy and Research Institute Co-development of Techniques’ funded by the Small and Medium Business Administration (R00046281).

References 1. http://en.wikipedia.org/wiki/IPTV 2. Park, S., Jeong, S., Hwang C.: Mobile IPTV Expanding the Value of IPTV. In: Seventh International Conference on Networking, pp.296-301. (2008) 3. Meng, S., Liu, L., Yin, J.: Mobile Scalable and Reliable IPTV Service Through Collaborative Request Dispatching. In: IEEE International Conference on Web Services, pp.179-186. (2010) 4. Hu, C., Hsu, Y., Hong, C., Hsu, S., Lin, Y., Hsu, C., Fang, T.: Home network management for IPTV service operations - A service provider perspective. In: Proceedings Network Operations and Management Symposium Workshops, pp.1-7. (2010) 5. Wagner, M., Kopilovic, I., Dohla, S., Farber, N., Fuchs, H.: Towards an RTP Profile for IPTV. IEEE Transactions on Broadcasting, Vol. 55, Issue 2, Part 2, 343--352 (2009) 6. Lee, E., Park, S., Lee, H.: Cross-Layer Design of Internet Group Management Protocol for Mobile IPTV Services in WiMAX. In: Fifth International Conference on Systems and Networks Communications, pp.294-299. (2010) 7. Yu, H., Yan, Y., Berger, M.: IPTV traffic management using topology-based hierarchical scheduling in carrier ethernet transport networks. In: Fourth International Conference on Communications and Networking in China, pp.1-5. (2009)

281

Lane Detection using Fuzzy C-Means Clustering Kwang-Baek Kim1, Doo Heon Song 2, Jae-Hyun Cho3 2

1 Dept. of Computer Engineering, Silla University, Busan, Korea Dept. of Computer Games, Yong-in SongDam University, Gyeonggi-do, Korea, 3 Dept. of Computer Engineering, Catholic University of Pusan, Busan, Korea [email protected], [email protected], [email protected]

Abstract. In general, road lane detection from a traffic surveillance camera is done by the analysis of geometric shapes of the road. Thus, Hough transform or B-snake technology is preferred to intelligent pattern matching or machine learning such as neural network. However, we insist that the feasibility of using intelligent technique in this area is quite undervalued. In this paper, we first divide the image into halves and use only the lower part in detection and binarize them by analyzing RGB channel. Then the boundary lines are extracted by applying 4-directional contour tracking algorithm and vectors with distance and angle values are extracted from those boundary lines to use as input for fuzzy C-means clustering algorithm. Clustered vectors form a straight line for road lanes by our method. In experiment, the proposed intelligent method is slightly slower than Hough transform but better in accuracy thus there is a room for intelligent method such as fuzzy C-means to solve this problem. Keywords: 4-directional contour tracking, Detection, Fuzzy C-means

1

Introduction

One of the frequent applications of image processing is ITS (Intelligent Transport System) using CC-TV [1]. One of the basic but important algorithms in such image analysis is lane detection. It is a basic building block of many traffic analysis applications such as lane change detection and collision detection. There has been a great deal of research into lane detection for operating intelligent vehicles. Hough transform that maps a line as an intersection point in the Hough area if it is translated to a coordinate system that has the relative rotation angle and distance from origin, i.e. the polar coordinate [2] and B-Snake technology [3] are two well known algorithms in this department. However, one may argue that these algorithms are basically based on the linear processing schemes. Recent trend in computing paradigm is to use MIMD with parallel processing in order to obtain maximum utility of resources such as cloud systems as an example. From that point of view, we argue that a system based on artificial intelligence such as fuzzy logic or neural network might be more efficient than linear systems in such environment even in low level ITS applications [4]. Such intelligent techniques have been mostly disregarded in problems such as lane detection since Hough transform or B-snake simply looks more efficient in linear processing environment.

282

2

Lane Detection using Fuzzy C-Means Clustering

We use 640 × 480 size 24 bit BMP format images as input. We first divide the image into halves and use only the lower part in detection and binarize them by analyzing RGB channel. The reason for using only lower half is that the upper part has only small portion of information about road lane thus it is inefficient considering the processing time. Road lanes are typically white or yellow thus we binarize them as such. Then the boundary lines are extracted by applying 4-directional contour tracking algorithm [5] and pixels having value less than 100 are removed as noise. Figure 1 summarizes these preprocessing steps.

Fig. 1. Image Preprocessing Steps

Input characteristic vectors are extracted by applying formula (1) having the distance and the angle for every 50 pixels

Vi =

(P {Dist Ang (P i

i

, Pk mod 50 + 50) k mod 50, Pk mod 50 + 50) k mod 50

(1)

where P is the coordinates of boundary line. Figure 2 shows the vector extraction process.

Fig. 2. Feature Vector Extraction

FCM clustering [6], originally designed in 1973 by Bezdek, is a classification algorithm that enumerates degrees of cluster membership for each data based on the distance from the center point. The membership degree is represented by a number between 0 and 1 and the cumulated sum of dataset is 1. In formula (2), Uik denotes the membership degree of kth data belongs to ith cluster. The center of jth cluster is computed by formula (3). Then the new membership degree Uik(r+1) is computed by

283

formula (4). The procedure is repeated until the difference between the new and old membership degree is smaller than the threshold value and as a result, input vectors are clustered as such. c

∑U

= 1, ∇k = 1,⋅ ⋅ ⋅, n

ik

i −1

(2)

n

∑ (U v ij =

ik

) m x kj

k =1 n

∑ (U

(3)

)m

ik

k =1

U ik

( r +1)

=

1  d ik r  ∑  r j =1  d jk c

   

2 / m −1

(4)

Figure 3 shows the plotting of input vector set into FCM.

Fig. 3. Vector Graph

Naturally, input vectors form a loosely coupled clusters as shown in Figure 3 since vectors from the same straight line(lane) have similar distance and angles. When we use FCM, we should provide the number of resultant clusters and it was 4 in this paper. Vectors are more clearly clustered after applying FCM as shown in Figure 4. Figure 5 shows the result of FCM comparing with the original image.

Fig. 4. Clustering Result

284

The original image (Figure 5 (a)) is divided into two parts and the lower half is binarized (Figure 5 (b)) and 4-directional contour tracking algorithm is applied (Figure 5 (c)) and the characteristic vectors are extracted (Figure 5 (d)) and FCM is applied to them as shown in Figure 5 (e).

Fig. 5. Original Image and Image after FCM

3

Experimental result and analysis

The software is written in Visual Studio 2005 on an IBM compatible PC with Intel Pentium-IV 3 GHz CPU and 1G RAM using images of 24 bit BMP format 640 × 480 in size.

285

Fig. 6. Experiment on the Number of Clusters

The main clustering algorithm FCM requires fixed initial number of clusters by nature. Although that is the main deficiency in this type of application, we use the number of clusters from 3 to 6 and compared them in Figure 6. Figure 6(b) has the number of cluster 3 and Figure 6 (c) has 4, Figure 6(d) has 5, and Figure 6(e) has 6. As shown in Figure 6, presetting the number of clusters is not a big problem in this application but for most cases, the two center lines are most important thus FCM using 4 as the number of clusters is most efficient in our experiment.

4

Conclusion

The common conception of applying intelligent algorithms such as fuzzy logic or neural network is finding an application that is not too much speed-sensitive and has less formal information like geometric characteristics. However, its possibility of adopting parallel processing schemes is another advantage for extending its usability. Our effort here using FCM in low-level ITS that has well deserved formal characteristics such as lane detection can only verify that the intelligent algorithm has been undervalued in such problems and it can be applied in real time as well if the environment allows (or requires) non-linear processing environment. Simple characteristic vectors extracted by image preprocessing are successfully clustered into a straight line without losing lane information even with partly disconnected real world road lanes in our experiment. However, the current method can only be

286

applied to straight lines and produces some amount of false positive objects and these problems will only stimulate us to develop the next research.

References 1. J. Palen, The need for surveillance in Intelligent Transportation Systems, Intellimotion, 6, 1 (1997) 2. B. Yu, A.K. Jain, Lane boundary detection using a multiresolution Hough transform, Proceedings of International Conference on Image Processing, 2, (1997) 3. Y. Wang, E. K. Teoh, D. Shen, Lane detection and tracking using B-Snake, Image and Vision Computing, 22, 4, (2004) 4. K. B. Kim, J. H. Cho, C. K. Kim, Recognition of Passports Using FCM-Based RBF Network, Lecture Notes in Artificial Intelligence, LNAI 3809, (2005) 5. R. Babuska, Fuzzy Modeling for Control, Kluwer Academic Publishers (1998) 6. W. S. Shin, D.H. Song, C.H. Lee, Vehicle Classification by Road Lane Detection and Model Fitting using a Surveillance Camera, International Journal of Information Processing Systems, 2, 1, (2006)

287

Performance Evaluation of HEVC Intra Prediction Algorithms Parallelized in a Mesh-based Multicore System Neungsoo Park1, Young-ho Park2 and Yungho Choi3, 2

1 Department of Computer Science and Engineering, Konkuk University, Department of Multimedia Science, Sookmyung Women’s University, Seoul, Korea 3 Department of Electrical Engineering, Konkuk University, 1 neungsoo @konkuk.ac.kr, [email protected], 3 [email protected] (Corresponding Author)

Abstract. For a high coding efficiency of HEVC, intra prediction algorithms are getting complicated. This necessitates multithreaded implementations for algorithms to shorten the coding time of HEVC. In this regard, this work parallelizes three intra prediction algorithms proposed for HEVC and evaluates their parallel performance using a mesh network-based multicore system, Tile64. Through this work, UDIP is shown to be best executed in Tile64 (almost 2 times faster than other algorithms). Keywords: Parallel, Intra, Prediction, Multicore system, Mesh network.

1

Introduction

Video compression technology plays a crucial role in implementing multimedia services by enabling to store and transmit video data with less resource. Such compression technologies include MPEG2, MPEG4, and H.264/AVC, which reduce data redundancy in subsequent video images. Since, these days, the need for higher image quality of multimedia services has been increased. To accommodate this need, JCTVC has been working on a new draft of a video compression technology, i.e., HEVC (High Efficiency Video Coding). This standard is designed for high resolution videos such as UDTV, which might not be efficiently compressed by previous compression technologies. To enhance compression efficiency, HEVC provides larger and more variable coding block sizes, adaptive loop filter, larger transform kernel size, etc. Despite its high compression efficiency, due to its coding complexity and tools’ dependencies, it might suffer from a long coding time. Therefore, it is important to implement HEVC in an environment of multicore systems to speed up its complicated coding processes. This paper explores the performance of intra prediction algorithms proposed for HEVC in a mesh-based multicore system. To do so, this work parallelizes each intra prediction algorithm by employing a wavefront scheme [1] which breaks off dependency chains among coding blocks and enables to execute an intra prediction

288

algorithm concurrently in many cores. Through evaluations, this work will provide an insight on how much parallelism each intra prediction algorithm has and how well each algorithm can exploit mesh-based multicore systems. The remainder of this paper is organized as follows: Section 2 briefly summarizes intra prediction algorithms to be explored. In Section 3, a performance evaluation on algorithms is given. In last, this work will be concluded in Section 4..

2

Intra Prediction Algorithms for HEVC

This work takes into three intra prediction algorithms, i.e., Unified Directional Intra Prediction (UDIP) [2,3], Combined Multi-parameter Intra Prediction (CIP) [2], Planar Intra Prediction (Planar) [3]. UDIP is an intra prediction algorithm which provides more various pixel prediction directions due to larger coding block sizes. CIP is similar to UDIP except for providing an additional low pass filtering-like process using multi parameters, thus better predicting natural images. Planar utilizes the most right-bottom pixel of the coding block to predict pixel values impacted by pixels of right and/or bottom sides, which suits for planar images. Since these algorithms behave differently in terms of data accessing pattern, communication among processing nodes, coding complexity, etc. in a mesh networkbased multicore system, it is important to analyze their processing characteristics. In the next section, an evaluation of a multithreaded performance in Tile64 is given to identify the algorithm best performing in a multicore system.

3

Performance Evaluation of Parallelized Intra Prediction Algorithms in a Mesh Network-based Multicore system.

This section evaluates the multithread performance of three intra prediction algorithms by parallelizing the algorithms implemented in TMuC1.0[4] using the pthread API of Tilera Tile64. Tile64 consists of 64 processing cores connected by a mesh network. The parallel algorithm employed for these intra prediction schemes is the wavefront algorithm [1] that enables a parallel processing at the level of coding blocks. For this multithread performance evaluation, each algorithm is executed in Tile64 under the following conditions: PeopleOnStreet(2560x1600) sequence, 64x64 maximum coding unit size, Intra-only configuration, 100 frames, QP=32. The results on the execution time and the speedup of each intra prediction algorithm are given in Figure 1. Speedup is obtained by dividing single-thread execution time of an algorithm by its multithread execution time. As shown in Fig. 1(a), CIP shows the highest speedup, which means the highest parallelism and the best exploitation of the given mesh network-based multicore system, Tile64 (up to 25% higher speedup than other algorithms). As the number of processing cores increases, the speed- up of algorithms is saturated around 8 cores. This indicates that, regardless of parallelism degree, the mesh network of Tile64 becomes saturated, which requires a careful optimization of communications among processing cores.

289

Fig. 1(b) shows parallel execution times of algorithms. As shown, UDIP can be executed up to two times faster than other algorithms. This means that UDIP has the lowest coding complexity and does not suffer from mesh network congestion significantly, while, although CIP shows the highest speed-up, its coding complexity is much larger than other algorithms, ending up with the longest execution time.

(a)

(b)

Speedup of intra prediction algorithms

Execution time of prediction algorithms (ms/frame)

Fig. 1. Parallel performance of intra prediction algorithms

4

Conclusion

This work parallelizes three intra prediction algorithms for a mesh network-based multicore system, Tile64, and, using these implementations, evaluates the parallel performance of intra prediction algorithms. Through this work, it is shown that, when intra prediction algorithms are parallelized and executed concurrently in multiprocessor systems, the network connecting processing cores might be a performance limitation by being saturated. Overall, UDIP can be best executed in Tile64 (almost 2 times faster than other algorithms). Additionally, in spite of the highest speedup, the high coding complexity of CIP increases its execution time. In this work, drawn is a conclusion that, in

290

selecting an algorithm for HEVC, taking into accounts both the coding complexity and the parallelizability of algorithms is necessary.

References 1. Zhuo Zhao, Ping Liang: Data partition for wavefront parallelization of H.264 video encoder. In: 2006 IEEE International Symposium on Circuits and Systems, pp.4-2672, 2006. 2. K. McCann, W. Han, I. Kim, J. Min, E. Alshina, T. Lee, J. Chen, V. Seregin, S. Lee, Y. Hong, M. Cheon, N. Shlyakhov.: Video Coding Technology proposal by Samsung. In: JCTVC-A124, Dresden, German, Apr. 3. K. Ugur, R. Andersson, A. Fuldseth.: Video Coding Technology Proposal by Tandberg, Nokia and Ericsson. In: JCTVC-A119, Dresden, German, Apr. 4. HEVC Reference Software TMuC1.0 [Online], Available: https://hevc.hhi.fraunhofer.de/svn/svn_HEVCSoftware/HM-1.0

291

A Proactive Anti-Phishing Tool Using Fuzzy Logic and RIPPER Data Mining Classification Algorithm Rosana J. Ferolin Department of Computer Engineering University of San Carlos, Cebu City, Philippines [email protected]

Abstract. Phishing emails are commonly used in fraud to illegally obtain confidential information mostly from clients of financial organizations. Phishing attacks are increasing in volume and have become more and more sophisticated which sometimes bypass the anti-phishing filter set by existing tools. This may be attributed to the fact that most of the approaches used by existing anti-phishing techniques are passive in form. Such approaches are passive in a way since they do not stop the source of the phishing emails or phishing websites but rather they simply classify and detect phishing emails.This paper presents a method for assessing and identifying phishing emails using Fuzzy Logic and the RIPPER Data Mining algorithm. In assessing the Phishing email, Fuzzy Logic linguistic descriptors are assigned to a range of values for each key phishing characteristic indicators. The membership function is designed for each Phishing email characteristic descriptor. The Data Mining RIPPER algorithm is used to characterize the Phishing emails and classify them based on both content-based and non-content based characteristics of Phishing emails. Furthermore, after the email has been assessed and classified as a Phishing email, the system proactively gets rid of the Phishing site or Phishing page by sending a notification to the Administrator of the host server that it is hosting a Phishing site which may result in the removal of the site.The initial results showed that the RIPPER algorithm achieved 85.4% for correctly classified Phishing emails and 14.6% for wrongly classified Phishing emails based on publicly available datasets from Phistank. After classifying the Phishing email, the system sends a notification to the System Administrator of the host server to indicate that it is hosting a Phishing site. Keywords: Data Mining, Fuzzy Logic, Phishing, URL, Classification

1

Introduction

Phishing is as an act of sending an e-mail to a user falsely claiming to be a legitimate business establishment in an attempt to scam or trick the user into surrendering private information that will be used for identity theft. It is a type of

292

network attack where the attacker creates a replica of an existing legitimate commercial website to deceive users to submit personal, financial, or confidential data to what to they think is their genuine business provider’s site. It is a security attack that involves obtaining private and classified data by presenting oneself as a reliable and genuine entity. The phishing e-mail then directs the user to visit a Web site that looks very similar to the Web site of the valid commercial establishment. Most of the time, Phishers take advantage of user’s trust on the appearance of a site by designing websites that are very similar to an authentic establishment’s website. The difference in appearance is most of the time unnoticeable by just merely looking at the site, where they are prompted to provide and update personal information such as passwords, bank account numbers, pin numbers, credit card numbers and other private information. This however, is a hoax and is set-up to steal the user’s information. Once the private information is obtained, the hacker can now login to the user’s account and process fraudulent transactions. This makes phishing an online fraud. The damage caused by phishing ranges from loss of access to email to significant and considerable financial loss. There were at least 67, 677 phishing attacks reported by the Anti-Phishing Working Group (APWG) in the last six months 0f 2010 [1]. The latest reports showed that most phishing attacks are “spear phishing” that aim the financial, business and payment sectors [2]. The number of phishing attacks and phishing sites is rapidly increasing and on the average, Sophos identifies 16,173 malicious web pages everyday [3]. Even though Web browsers (i.e. Mozilla Firefox, Internet Explorer, Opera, Google Chrome, etc.) provide add-on tool for blocking phishing emails and phishing sites, Phishers still manage to override these security mechanisms. Phishing has become more and more complicated that Phishers can bypass the filter set by current anti- phishing techniques. The rapidly increasing number of phishing attacks suggests that it is therefore difficult to find a single logical procedure to detect phishing emails and that existing anti-phishing tools are not sufficient. This may be attributed to the mostly passive approach of anti-phishing techniques. The approaches are passive since they do not stop the source of the phishing emails rather they simply classify and detect phishing emails [4]. A proactive approach to minimizing phishing has been conducted where the system removes a phishing page from the host server rather than just filtering email and flagging suspected messages as spam [4]. The study in [4] however, assumes that emails have already been classified as a phishing email or legitimate email. The study has ignored the phishing email classification and was more concerned with how to deal with the Phisher once a phishing email has been detected. This study proposes to develop an anti-phishing tool that combines phishing email classification and the proactive approach to stop the Phishers from its source as a result of the classification. This will be an improvement of the of the existing study in [4] where classification was ignored and that the proactive approach is based on the assumption that emails have been already identified as phishing emails. The study will take into consideration different email features in classifying phishing emails. The aim of this study is to use Fuzzy Logic and Data Mining Classification algorithm in classifying

293

and detecting phishing emails. In addition, once an email is classified as a phishing email, the system will proactively stop phishing at its source by tracing the website and informing the host server of the of the phishing activity.

2

Related Studies

Most anti-phishing tools employ email filtering techniques to classify legitimate emails and suspected spam in the mail inbox. The user is left to decide whether to open or discard such emails. If no anti-phishing tool is installed or the user has not updated the anti-phishing program, then there is no layer of protection. This is referred to as passive anti-phishing [4]. It is because the approach only locally protects the user from a phishing attack but does not make any effort to stop or remove the Phisher at the source. The Phisher then continues with the phishing operation to further increase its victims. While there are several email filters, browser tools, anti-spyware and anti-virus software, very few research efforts have been entirely focused to protect online users from phishing attacks in the past. Existing anti-phishing and anti-spam techniques suffer from one or more limitations and they are not 100% effective at stopping all spam and phishing attacks [9]. Phishers are able to find ways to bypass existing rule-based and statistical based filters without much difficulty. Major e-mail service providers such as Yahoo, Hotmail, Gmail, and AOL filter all incoming emails separating them into Inbox (legitimate email) and junk (illegitimate email) email folders. However, these email service providers do not actually attempt to remove the phishing page associated with the illegitimate email. Furthermore, Phishers have readily available tools to bypass such spam filters [5]. We refer to this as a passive anti-phishing approach. This is because the approach only attempts to locally protect an individual from a phishing attack, but does not actively make any effort to remove or shut down the Phisher at the source. In effect, the Phisher is free to continue with the fraudulent operation and can potentially accrue further victims. There are different existing phishing detection approaches. These approaches can be further classified as 1) content-based approaches that use site contents to detect phishing, 2) non-content based approaches that do not use the content of the site to identify if the email is an authentic or phishing email and 3) Visual based approaches that identify phishing using the similarity of known sites through visual inspection [5].

294

A. Content Based Approach. In content based approach, phishing attacks are detected by examining site contents. Features used in this approach include keywords, spelling errors, links, password fields, embedded links, etc. along with URL and host based features [5]. Google’s anti-phishing filter detects phishing and malware by examining page URL, page rank, WHOIS information and contents of a page including HTML, javascript, images, iframe, etc. [5]. The classifier is constantly updated to accommodate new phishing sites to cope up with the latest techniques in phishing attacks. In this approach the classifier may have higher accuracy but the result is not real-time. It is used offline since it takes longer to detect the Phishing. Several researchers have explored different approaches such as fingerprinting, principal component analysis of images, heuristic approaches and fuzzy logic among others and fuzzy logic based approaches to identify phishing sites. Our approach uses Fuzzy Logic language descriptors with a range of values for each identified phishing characteristic specifically spelling errors, keywords and embedded links. The membership function for each characteristic derived as is used to assess the probability that the email is a phishing email. B. Non-Content Based Approach. Non-content based approaches are primarily based on URL and host information classification. URLs are commonly classified based on features such as URL address length and presence of special characters. Moreover, host features of URL such as IP address, site owner, DNS properties and geographical properties are also used in the classification of Phishing emails [5]. The success rate is between 95% - 99% even in real-time processing [15].

3

Methodology

3.1

System Flow

This section describes the overall approach of the system in assessing, detecting and classifying Phishing emails. The process for notifying the hosting site or the sending site of the Phishing email is also included as we ll as the possible removal process. At the start, the system assesses the risk of the email using Fuzzy Logic. It then classifies the email as Phishing or legitimate email. The classification makes use of the data mining RIPPER algorithm. If the system detects that it is a phishing email, it gets the URL of the Phishing email. The host server’s IP address, host server location and the contact information of the System Administrator. A notification is sent to the System Administrator of the host server informing that a phishing page is hosted by the server. The System Administrator proceeds with the removal of the Phishing page.

295

Figure 1. Overall system approach

3.2

Detecting and Classifying Phishing Email

A. Main Characteristics of Phishing Emails. The proposed methodology will apply fuzzy logic and data mining algorithms to classify phishing emails based on two classification approaches such as content-based approach and non-content based approach. Specific categories or criteria are selected for each approach. The components or selected features are then identified for each category. The list of the classification approaches with the identified criteria and specific features is listed in the table below. The list will be used as basis for in the simulation and determination of phishing emails.

296

The main characteristics of phishing emails are listed in Table 1. Table I. Characteristics and stages of the components of phishing emails

Classification Approach Category/Criteria Non-content Based URL Approach

Content-based Approach Email Message

Overall Weight

3.3

Component Stage/Layer IP URL Stage 1 Redirect URL Non-matching URL Crawler URL Long URl address Weight = 0.5 URL prefix/suffix Spelling Errors Stage 2 Keywords Weight = 0.5 Embedded links 1.0

Use of Fuzzy Logic and RIPPER Data Mining Algorithm

The approach is to apply fuzzy logic and RIPPER data mining algorithm to assess phishing email based on the 9 identified characteristics or components. The essential advantage offered by fuzzy logic techniques is the use of linguistic variables to represent key phishing characteristic or indicators in relating phishing email probability. A. Fuzzification. During fuzzification, linguistic descriptors such as High, Low, Medium, for example, are assigned to a range of values for each key phishing characteristic indicators. Valid ranges of the inputs are considered and divided into classes, or fuzzy sets [7]. For example, redirect URL can range from „low‟ to „high‟ with other values in between. The degree of belongingness of the values of the variables to any selected class is called the degree of membership; Membership function is designed for each Phishing characteristic indicator. Each point in the input space is mapped to a membership value between [0, 1]. For each input the values ranges from 0 to 6 while for output, the value ranges from 0 to 100. A plot of the membership function of the (Redirect URL) which is one of the phishing characteristic and the linguistics descriptors is illustrated in Figure 1 below. The same approach is used to calibrate the other 8 key phishing characteristic indicators.

297

Figure 1. An example of the linguistic descriptors that will be used to represent one of the key phishing characteristic indicators and a plot of the fuzzy membership functions

B. Rule Generation using the RIPPER classification algorithm. This study proposes to use the RIPPER algorithm to learn the relationships of selected different phishing features. Such classification algorithm is selected to learn the relationships of the different phishing features. Fuzzy rules are represented as if-then statement which is based on the phishing email probability with the different phishing characteristics or indicators. Classification is done using WEKA.

Figure 2. Determination of the Phishing email risk rate

298

C. Rule Outputs. The outputs of all discovered rules will be unified. Combining the membership function of all the rules ensuing consequent previously scaled into single fuzzy sets or outputs. D. Deffuzification. Deffuzification is the process of producing a measurable result in fuzzy logic given the fuzzy sets and membership degrees. It is a process in fuzzy logic where valuable data is produced from fuzzy data. This process transforms a fuzzy output of a fuzzy inference system into a crisp output [12]. Fuzzification facilitates in evaluating the rules, but the final output has to be a crisp number. The input for the defuzzification process is the collective fuzzy set and the output is a number. A useful defuzzification technique is the center of gravity. A typical set membership function has the graph of a triangle. The first step of defuzzification normally removes parts of the graph to form a trapezoid. The trapezoids are then superimposed one after the other to form a single geometric shape. The centroids which is called fuzzy centroid, is calculated. The x coordinate of the centroid is the defuzzified value. The output is the Phishing email risk rate and is defined in fuzzy sets as “Fraudulent” to “Genuine or Legitimate”. 3.4

Fuzzy Data Mining Inference Rules

A. Rule Base for Stage 1 or Layer 1. There are six input parameters for the rule base and it has one output. It contains all the if-then rules of the system. In each rule base, every component is assumed to be one of three values (based on the linguistic descriptor) and each criterion has six components. Hence rule base 1 contains 36 = 729 entries. The output of rule base 1 is one of the phishing email rate fuzzy sets (Genuine, Suspicious or Fraud) representing URL & Domain Identity criteria phishing risk rate. A sample of the structure and the entries of the rule base 1 for layer 1 are sho wn in Table IV. The system structure for URL & Domain Identity criteria is the joining of its five components, which produces the URL & Domain Identity criteria (Layer one). Table 2. Sample of the rule base stage 1 entries for the URL Domain and Entity Criteria Rule #

1 2 3 4 5 6 7

IP URL

Crawler Redirect NonURL URL matching URL

Low Low Low Low Low Low Low Low Low Low Low Moderate Moderate Low Moderate Low Low Low Low Moderate Moderate Moderate Moderate High High High High High

299

Long URL URL URL Domain address Prefix/suffix Entity & Criteria Low Low Moderate Low Moderate High Moderate

Low Moderate Moderate Moderate High High Moderate

Valid/Genuine Valid/Genuine Suspicious Suspicious Suspicious Fraud Fraud

B. Rule Base for Stage 2 Table 3. Sample of the rule base stage 2 entries for Email Content Domain Rule # Spelling Errors Keywords 1 2 3 4 5 6 7

Low Low High Low High Moderate Moderate

Embedded Links

Low Moderate High Low Moderate Low Moderate

Moderate Moderate High Low Moderate Moderate Low

Email Content Domain Genuine Suspicious Fraud Genuine Fraud Suspicious Suspicious

3.5 Locating the Host Server of the Phishing Page WHOIS is a protocol used to find information about networks, domains and hosts. The WHOIS query is used to locate the host server of a phishing page. WHOIS is a query/response protocol that is widely used for querying an official database. The WHOIS database contains IP addresses, autonomous system numbers, organizations or customers that are associated with these resources, and related Points of Contact on the Internet [4]. A WHOIS search will provide information regarding a domain name, such as example.com. It may include information, such as domain ownership, where and when registered, expiration date, and the name servers assigned to the domain. A WHOIS server listens on Transmission Control Protocol port 43 for requests of the host server and related contact information sent through web-based referrals [4]. Its connection is closed as soon as the output is finished. The closed connection is the indication to the client that the response has been received. The system runs the WHOIS query on the URL that is contained in the Phishing email. 3.6 Removal of the Phishing Page Upon receiving the notification of the phishing page‟s existence on the host server, the hosting administrator will then test the legitimacy of the phishing link and its validity. Once the Administrator confirms the phishing page, the infected or hacked website will be shut down immediately to protect Internet users from further phishing. The host Administrator then notifies the website owner about the existence of the phishing page within their website. As soon as the phishing page is removed, if no notification has been sent, the proposed system will periodically check for evidence that it has been removed. This technique assumes that website owner and host Administrator are absolutely unaware of the presence of the phishing page within their website or server until our

300

technique notifies them. This means Phishers are taking control of the legitimate website to upload their phishing page. By doing so phishing pages are able to bypass anti- virus securities installed on the user’s computer.

4

Results

Publicly available datasets from Phistank were used for simulation. There are two stages in determining the fuzzy data mining inference rules. 1000 sample instances are used from the Phistank archive. For rule base 1, there are 6 identified Phishing email characteristics based on the non-content based approach. The assigned weight is 0.5. For rule base 2, there are 3 identified characteristics of Phishing emails based on the content-based approach. The assigned weight is 0.5. The email rating is computed as 0.5 * URL and Domain Entity crisp (rule base 1) + 0.5 * Email Content Domain crisp (rule base 2).

Figure 3. Fuzzy logic inference overall system use in assessing the Phishing email risk rate

The study utilizes the RIPPER data mining algorithm. The RIPPER algorithm uses separate and conquer approach. It is considered an inductive rule learner that builds a set of rules that identify the classes while minimizing the amount of error. The error is determined by the number of training examples that are misclassified by the rules. The prediction accuracy is recorded in Table 4. Table 4. Results generated from the WEKA classifier using RIPPER algorithm applied to classify Phishing emails

10 fold cross validation URL Domain and Entity Criteria Email Content Domain Number of rules 12 Correctly classified 85.4% Incorrectly classified 14.6% Number of samples/instances 1000

Validation Mode Attributes

301

The initial results showed that URL and Entity Domain and the Email Content Domain are important criteria for identify and detecting Phishing emails. If one of them is “Valid or Genuine”, it will likely follow that the email is a legitimate email. The same is true if both of the criteria are “Valid or Genuine”. Likewise, if the criteria are “Fraud”, the email is considered as a Phishing email. The results regarding how many Phishing sites were removed as a result of the notifications sent after the email is classified as a Phishing email is yet to be established. Moreover, the final results for the input combinations shown in Table 1 and Table 2 is still to be determined. At this stage, data gathering is still on-going.

5

Conclusion

URL and Entity Domain as well as Email Content Domain are two important and significant Phishing criteria. If one of the criteria is “Valid or Genuine”, it will likely follow that the email is a legitimate email. The same is true if both of the criteria are “Valid or Genuine”. Likewise, if the criteria are “Fraud”, the email is considered as a Phishing email. It should be noted, however, that even if some of the Phishing email characteristics or stage is present, it does not automatically mean that the email is a Phishing email. The initial objective is to assess the risk of the email in the archive data using fuzzy logic and the RIPPER classification algorithm. Several characteristics were identified and major rules that were determined along the study were used in the fuzzy rule engine. Moreover, the number of Phishing sites removed as a result of sending a notification to the host server as soon as the email is classified as Phishy should be monitored to determine the effectiveness of the study. There may be instances where the host server will not remove the Phishing site even if it has received notifications. This case is beyond the context of the study as this requires administrative sanctions or inclusion of the site in the Blacklist. The study however was able to prove that fuzzy logic and data mining with the use of the RIPPER algorithm is in a way sufficient in assessing the risk of a Phishing email and classifying the email as such, thereby resulting in the issuance of notification to the host server for removal of the Phishing page.

References [1] A.-P. W. Group, “Global phishing survey: Domain name use and trends in 2h2010,”http://www.antiphishing.org/reports/APWG, GlobalPhishingSurvey 2H2010.pdf. [2] A. P. W. Group, “Phishing activity trends report,” 2009, http://www.antiphishing.org/reports/apwg reportQ42009.pdf. [3] Anti Phishing Working Group, “Phishing Activity Trend Report”, Jan-March 2008

302

[4] Shah, R.; Trevathan, J.; Read, W.; Ghodosi, H.,”A Proactive Approach to Preventing Phishing Attacks Using the Pshark Model”, IEEE Sixth International Conference on Information Technology: New Generations, March 2009, pp. 915 - 921 [5] Afroz, S.; Greenstadt, R., “PhishZoo: Detecting Phishing Websites by Looking at Them”, IEEE Fifth International Conference on Semantic Computing (ICSC), 2011, pp. 368-375 [6] Sophos, “Security Threat Report”, July – 2008 [7] Maher Aburrous, M. A. Hossain, Keshav Dahal, Fadi Thabatah, “Modelling Intelligent Phishing Detection System for e-Banking using Fuzzy Data Mining”, IEEE International Conference on CyberWorlds, 2009, pp265 - 272 [8] S. M. Bridges and R. B. Vaughn, “fuzzy data mining and genetic algorithms applied to intrusion detection,” Department of Computer Science Mississippi State University, White Paper, 2001.

303

[9] Rokach, Lior; Oded Maimon (2008). Data mining with decision trees: theory and applications. World Scientific Publishing. ISBN 978-9812771711. [10] L. James, “Phishing Exposed,” Tech Target Article sponsored by: Sunbelt software, searchexchange.com, 2006 [11] http://www.phishtank.com/phish_archive.php [12] J. Milletaty, “Technical Trends in Phishing Attacks”, US-CERT, http://www.uscert.gov/reading_room/ [13] T. Sharif, “Phishing Filter in IE7,”http://blogs.msdn.com/ie/archive/2005/09/09/463204.aspx September 9, 2006. [14] L. Wood, “Document Object Model Level 1 Specification,” http://www.w3.org, 2005. [15] J. Ma, L. K. Saul, S. Savage, and G. M. Voelker, “Identifying Suspicious URLS: An Application of Large-scale Online learning,” in ICML ’09:Proceedings of the International Conference on Machine Learning,2009, pp. 681–688.

304

Real Time Inverse Kinematics Robot Gait Sequencing Using Self Organizing Maps Chris Jordan G. Aliac Intelligent Systems Laboratory, Cebu Institute of Technology, Cebu City, 6000 Philippines [email protected]

Abstract. Real-time control of end-effector of humanoid robots in Cartesian coordinates requires computationally efficient solutions, especially when the complexity of analytic solutions becomes too high along with the number of degrees-of-freedom. In this study, we solve the inverse kinematics problem for dynamic gait sequencing combining mathematical modeling and neural networks and applying this to an actual Biped Robot. Forward kinematics transforms robotic configurations in joint space, into end- effector locations in Cartesian space. Pseudo-learning is done by associating the Cartesian space into the joint space until convergence. Our learning strategy is based on the topology-preserving property of the Extended Kohonen Model both in the competitive and output layers, even in rather high-dimensional spaces. The virtual output angles of the self-organizing maps are then mapped to the corresponding servomotor actuation values before being transmitted to the robot for control. The theoretical and actual results are illustrated with an integrated 3D simulation system and a 10 degree-of-freedom biped robot employing Bluetooth communication technology and Microsoft SAPI speech software in task issuance, respectively.

1

Introduction

The analytic approach in inverse-kinematics is first to derive the forwardkinematics equations p = f(θ), where p is a vector of Cartesian coordinates of the robot's end-effector, and θ is a vector of joint variables, and then solve for the inverse-kinematics equations θ = f-1(p). Finding such linear equations, if f(θ) is not uniquely invertible at all, requires much mathematical experience and precise knowledge of the robot's mechanical structure as well. No general solutions exist, thus, the term pseudoinverse functions apply to the inverse-kinematics equations. In addition, the computational complexity greatly increases with the robot's degree-of-freedom, as is the case with the 10 degree-of- freedom biped robot used in this study.

305

Fig. 1. Biped robot in this study

Prerequisite to the process of learning inverse kinematics is a framework for deriving training data. Since in this study the purpose of solving inverse kinematics is dynamic gait sequencing, gaits are de- fined based on kinematic trajectories in Cartesian space, and it is ther e- fore vital to have a mapping θ → p. Before transformations from joint to Cartesian spaces are realized, a mathematical model of the robot must be laid out. This is with the usage of DH Parameters di, ai, αi, and θi, describing relationships between links of a kinematic chain, where θi are variables and are initialized into “home position” values. Table 1. -1 : DH Parameters for the robot in study

306

Fig. 2. Mathematical Model

With DH parameters, the homogenous frame transforms frame Fi in the kinematic chain takes the form

i-1Pi for every link

The forward kinematic equations for the end-effector locations are finally derived as a product of successive homogenous frame tran s- forms in the link 0 0 1 n-1 P n, P n = P 1 P 2 ...

307

where

In the case of our biped robot, the values for the elements of the homogenous frame transforms are as follows: nLx = -cθ L2 c(θ L3 + θ L4 + θ L5) cθ L6 + sθ L2 sθ L6 nLy = -s(θ L3 + θ L4 + θ L5)cθ L6 nLz = sθ L2 c(θ L3 + θ L4 + θ L5) cθ L6 + cθ L2 sθ L6 sLx = cθ L2 c(θ L3 + θ L4 + θ L5) sθ L6 + sθ L2 cθ L6 sLy = s(θ L3 + θ L4 + θ L5) sθ L6 sLz = -sθ L2 c(θ L3 + θ L4 + θ L5) sθ L6 + cθ L2 cθ L6 t Lx = -cθ L2 s(θ L3 + θ L4 + θ L5) t Ly = c(θ L3 + θ L4 + θ L5) t Lz = sθ L2 s(θ L3 + θ L4 + θ L5) p Lx = -[a L6 c(θ L3 + θ L4 + θ L5) cθ L6 + a L5 c(θ L3 + θ L4 + θ L5) + a L4 c(θ L3 + θ L4) + a L3cθ L3 + a L2 + dL6 s(θ L3 + θ L4 + θ L5)] cθ L2 – a L1 + (a L6 cθ L6 + dL5 + dL4 + dL3) sθ L2 p Ly = -[a L6 s(θ L3 + θ L4 + θ L5) cθ L6 + a L5 s(θ L3 + θ L4 + θ L5) + a L4 s(θ L3 + θ L4) + a L3 sθ L3] + dL6 c(θ L3 + θ L4 + θ L5) + dL2 p Lz = [a L6 c(θ L3 + θ L4 + θ L5) cθ L6 + a L5 c(θ L3 + θ L4 + θ L5) + a L4 c(θ L3 + θ L4) + a L3 cθ L3 + a L2 + dL6 s(θ L3 + θ L4 + θ L5)] sθ L2 + (a L6 sθ L6 + dL5 + dL4 + dL3 ) cθ L2 nRx = cθ R2 c(θ R3 + θR4 + θ R5) cθ R6 + sθ R2 sθR6 nRx = s(θ R3 + θ R4 + θ R5) cθ R6 nRz = -[sθ R2 c(θ R3 + θ R 4 + θ R5) cθ R6 + cθR2 sθ R6] sRx = -[cθ R2 c(θ R3 + θ R4 + θ R5) sθ R6 + sθ R2 cθ R6] sRy = -s(θR3 + θ R4 + θ R5) sθR6 sRz = sθ R2 c(θ R3 + θ R4 + θ R5) sθ R6 - cθ R2 cθ R6 tRx = -cθR2 s(θ R3 + θ R4 + θR5 ) tRy = c(θR3 + θ R4 + θ R5) tRz = sθ R2 s(θ R3 + θ R4 + θR5 )

308

pRx = [a R6 c(θ R3 + θ R4 + θ R5) cθ R6 + a R5 c(θ R3 + θ R4 + θ R5) + aR4 c(θR3 + θR4) +a R3 cθ R3 + a R2 - dR6 s(θ R3 + θ R4 + θR5)] cθ R2 + a R1 + (dR5 + dR4 + dR3 - a R6 sθ R6) sθR2 pRy = a R6 s(θ R3 + θ R4 + θ R5) cθ R6 + aR5 s(θR3 + θ R4 + θ R5) + aR4 s(θ R3 + θR4) + a R3 sθR3 + dR6 c(θ R3 + θ R4 + θR5) + dR2 pRz = -[a R6 c(θ R3 + θ R4 + θR5) cθ R6 + a R5 c(θ R3 + θ R4 + θ R5) + a R4 c(θ R3 + θR4) + a R3 cθ R3 + aR2 - dR6 s(θ R3 + θ R4 + θ R5)] sθ R2 + (dR5 + dR4 + dR3 – aR6 sθ R6) cθ R2

In our approach, only the position vector p is relevant to inverse kinematics pseudo-learning, wherein p = f([θ2 θ3 θ4 θ5 θ6]) is a function of 5 joint variables for each leg, disregarding θ1, which is a constant. As training data, the Cartesian coordinates p = (px, py, pz) are the input signals and the joint values θ = [θ2 θ3 θ4 θ5 θ6] are the associated target output values. Ranges for the actuation values of servomotors are exactly [1, 254], while ranges for the angular values in the mathematical model vary. Offline control of servomotors require that virtual joint values be mapped into its corresponding actuator value, since the subject of learning are the virtual joint values. The following generic one-to-one map- ping are proposed, wherein four parameters, namely home virtual angle θv, home, home actual angle θa, home, maximum actual angle θa, max, and minimum actual angle θi, are defined. There are two types of ma p- ping, the parallel and the anti-parallel mapping, wherein the former, both values in the actual and virtual angle ranges increase as the actuation goes to the same direction, while in the latter, both values in the actual and virtual angle ranges increase as the actuation goes to the opposite direction. Derivation is based on discrete mathematics.

2

Inverse Kinematics Pseudolearning

The Extended Kohonen Model consists of an input layer, competitive layer, and an output layer, and its mode of learning is supervised learning due to the introduction of target output values together with input signals during training. In this study, two self-organizing maps were used, one for each leg, of which the rationale is to be able to segregate the learning process for each leg, since the intersection of the task spaces of each leg is not the null space, the implication of which is an event we term as “confusion”. For illustration, a Cartesian point reached by the left leg can also be reached by the right leg, but the joint values are different. Upon introduction of these inputs, the single self- organizing map will memorize output values that are conflicting, and the result during deployment will be an ill-posed kinematic chain. The dimensions of the task space vary according to the DH parameters. Thus the number of neural units in the three dimensional neural networks must be dynamic and computed depending on the set of training data G = {[pxpy pz θ2 θ3 θ4 θ5 θ6]} since the set training data is composed of Cartesian coordinates and corresponding joint angles in the task space. A constant r is also defined, which is the number of neurons assigned to subspaces of size r3 distance units. The

309

task space is represented by two Cartesian coordinates pmin = (px,min, py,min, pz,min) and pmax = (px,max, py,max, pz,max).

Fig. 3. -1: Task space

From pmin and pmax, the dimensions b, d, and h of the neural networks are computed. b = ceil( (|floor(xmin – a)| + floor(xmax + a)|) / r) d = ceil( (|floor(ymin – a)| + floor(ymax + a)|) / r) h = ceil( (|floor(zmin – a)| + floor(zmax + a)|) / r) where |n| computes the absolute value of n, ceil and floor are the usual discrete integer functions.

Fig. 4. -2: The neural unit - subspace association ratio r

Parameters to the training phase include initial learning rate ή0, ini- tial

310

neighborhood size δ0, input bias xinp, output bias xout, and maxi- mum number of epochs T. The learning algorithms are as enlisted below: 1. Initialize node connection weights with random 2. values in the range (0,1). 3. For all training data g = [px py pz θ2 θ3 θ4 θ5 θ6 s], where s indicates which leg is in scrutiny, locate the neuron hi,j,k responsible for the input signal p = [px py pz] and initialize output weights with θ = [θ2 θ3 θ4 θ5 θ6] where, i = floor( ( px – floor(px,min) ) / r ) j = floor( ( py – floor(py,min) ) / r ) k = floor( ( pz – floor(pz,min) ) / r ) 3. For each epoch t: 0 → T 3.1. For p: 0 → n { ' 2 1. Find winning neuron r = argmin i, j, k || p - w || t t-1 t-1 2. For all neuron r, update wr = wr + ήtdrr'(p – wr) t ' t-1 t-1 ' r 3. For all neuron r, update θ r = θ'r + ήtdrr'(θ – θ ) } 3.2. Update learning rate ήt = (1 - t/T)ήt-1 3.3. Update neighborhood size δt = (1 – t/T)δ t-1, where drr' = e ^ -(|| p - w ||2 / 2δt 2)) and n is the number of training data g. Average input and output weights error per epoch t are inspected during training, since graphs asymptotic to zero could possibly indicate convergence on competitive and/or output layers. Computations are as follows. n 2 εinput,k = Σi || pi – wi || / n n 2 εoutput,k = Σ i || θi – θ'i || / n

3

Simulation Systems

Three inter-related simulation systems were designed for this study namely, the Gaits Designer, the Humanoid Robot Training Simulator, and the Intelligent Humanoid Robot Behaviour, implemented in the object-oriented language C#.

311

In the training phase, gait phases are created, and the joint values are input into the Gaits Designer. These gait phases can be stored as a .gts file for later manipulation or as a .trd file for training. The Humanoid Robot Training Simulator then loads the .trd file and starts training after learning parameters have been set. The resulting weights can then finally be stored as a .kts file and loaded in the Intelligent Humanoid Robot Behaviour for deployment. Voice commands issued are recognized by the SAPI voice tool, and corresponding dynamic gait sequences are invoked. After mapping virtual to actual values, actuation values for each servomotor are transmitted to the biped robot. All three simulation systems render three dimensional viewing of the kinematic chain, self-organizing maps, task space samples, and other entities, using a transform matrix P for display of 3D Cartesian coordinates.

Fig. 5.

Figure 3-1: Training Phase Subsystems Architecture

312

Fig. 6. -2: Deployment Phase Subsystems Architecture

Moreover, to display 3D coordinates in 2D graphics, an angle φ, aside from display parameters α, β, and γ, is employed, where x' = x + ycosφ and y' = z + ysinφ, such that (x, y, z) is a 3D coordinate and (x', y') is a 2D coordinate.

References 1. Alviric Alcarez, Jasper Baranda, Leo Angelo Lumanog (2009). Bluetooth Controlled Biped Robot. Cebu Institute of Technology. 2. Maravillas, E.A. Real-Time Vision-Based Robot Manipulator Control Using Self Organizing Maps. Cebu Institute of Technology. 3. Bachar, Yariv (2004). Developing Controllers for Biped Humanoid Locomotion. University of Edinburgh: School of Informatics. 4. Chung-Hsien Kuo, Yao-Chun Hung, En-Chou Lin, Jian-Chih Chen, Mike Lee, Chuan-Chia Sun. Autonomous Humanoid Soccer Robot: Taiwan 101 V1. Graduate Institute of Medical Mechatronics, Chang Gung University, Taiwan. Aaron D’Souza, Sethu Vijaya- kumar and Stefan Schaal. Learning Inverse Kinematics. Computer Science and Neuros- cience, HNB103, Univ. of Southern California, Los Angeles, CA 90089-2520, Kawato 5. Dynamic Brain Project (ERATO/JST), 2-2 Hikaridai, Seika-cho, Soraku-gun, 619-02 Kyoto, Japan. 6. C. Versino, L.M. Gambardella. Learning Fine Motion in Robotics: Design and Experiments. IDSIA, Corso Elvezia 36, 6900 Lugano,Switzerland. 7. Sears and Zemansky’s University Physics. Young, Hugh D. 2005 8. Robot Modeling and Kinematics. Manseur, Rachid. 2003 9. http://en.wikipedia.org/wiki/Selforganizing_ map 10. http://en.wikipedia.org/wiki/Denavit- Hartenberg_Parameters 11. http://davis.wpi.edu/~matt/courses/soms/

313

Agent-based Safety Support System for Collision Avoidance in Dilemma Zones Eunyoung Lee1, Hyung-Ju Cho2, Kiyeol Ryu2 1Graduate

School of Information and Communication, Ajou University, South Korea of Information and Computer Engineering, Ajou University, South Korea

2Department

{nescafe,hjcho,kryu}@ajou.ac.kr

Abstract. This study proposes an agent-based intersection system to prevent accidents and improve the safety at intersections. In particular, the proposed algorithm is targeted at a collision-avoidance service for use at the dilemma zone of signalized intersections. To collect and deliver information, the agent technology is applied to both vehicles and intersections. Based on this information, the system analyzes and determines the possibility of accidents using an algorithm. Next, it takes appropriate actions depending on the particular situation at hand. The agent-based intersection system solves priority and information conflicts through the interaction and cooperation among agents. The system also supports the prevention of accidents through risk prediction based on information regarding driver tendencies. In addition, the system provides warnings and controls subsystems, such as traffic signals and vehicle brake systems, based on the decisions made. Finally, simulations show the status of an intersection before and after the agent technology is applied. Keywords: Agent System, Collision Warning and Avoidance, Dilemma Zone, Intersection System

1

Introduction

According to the 2011 traffic accident statistics from the Korean National Police Agency, 101,547 accidents occurred at intersections in Korea, corresponding to 44.7% of all reported traffic accidents nationwide in 2010 [1]. Depending on the country and region, 30 to 60% of all crashes causing injury or fatalities take place at intersections [2]. This data shows that it is very important to improve an intersection system for transportation safety. Therefore, many studies have been conducted to improve the safety at intersections around the world. The goal of this study is to develop a support system for collision avoidance at intersections by utilizing an agentbased technology. An agent-based technology is suitable to traffic domains that are distributed geographically in dynamic environments, where intersections are the common problem domains. An agent-based technology can vastly improve the design and analysis of problem domains, especially when the problem domain is distributed geographically, and when the subsystems are in a dynamic environment and need to interact with each 1

314

other in a very flexible manner [3]. Agents communicate with other agents in the system to achieve a common goal. They can also perceive their environment and react to environmental changes in a timely fashion. Applying these characteristics of an agent-based technology to a transportation system, agent-based transportation systems allow distributed subsystems to collaborate with each other and manage traffic controls based on real-time traffic situations [4]. To apply an agent-based technology to an intersection system, the types of accidents that occur at intersections were first identified. It was found that 20% of all traffic accidents are collisions in intersections [5], and in 2010, 11% of all accidents were related to traffic signals [6]. While high fatality rates during light traffic are caused by traffic-signal violations, the crash rate during heavy traffic appears to be higher in the dilemma zone. Therefore, this study focuses on collision avoidance services in dilemma zones at signal intersections. Recent studies have focused on vehicle detection and warning systems in the dilemma zone using V2V (Vehicle-to-Vehicle) and V2I (Vehicle-toInfrastructure) communication technologies [5]. However, little attention has been paid to priority and information conflicts. Thus, this study proposes an agent-based intersection system, the primary contributions of which can be summarized as follows:  The proposed system improves intersection safety through interactions and cooperation among agents. The agent-based intersection system solves priority and information conflicts.  This system allows a targeted service to be available to each vehicle during a dangerous situation.  This system reevaluates the risk of accidents based on the driver’s collision avoidance response after receiving a warning.  The proposed algorithms helps to avoid rear-end and right-angle collisions in the dilemma zone. The rest of this paper is structured as follows: Section 2 surveys the related work. Section 3 describes the agent-based system for intersection safety. Section 4 details the algorithm used for collision avoidance. Section 5 presents simple prototypes of the proposed system. Finally, Section 6 provides some concluding remarks regarding this research.

2

Related work

There have been many studies and technological improvements in the area of transportation safety, such as an ITS (Intelligent Transport System), Telematics, and a Cooperative ITS. In recent years, various safety services, including intersection collision warning and avoidance, have been realized. Table 1 shows some existing projects related to intersection safety. These projects are focused on real-time vehicle detection and warning using ubiquitous computing technologies such as USN (Ubiquitous Sensor Network), V2V, and V2I.

2

315

A cooperative intersection collision warning and avoidance system collects data from vehicles and roadside sensors, and performs real-time detection and warning using communication technologies [7]. Table 1. Projects related to intersection safety Project Name Region Applications CICAS USA Right of way (2005~now) Red light violation [14] Stop sign assistance Right turn assistance Smart way Japan Traffic light scenario (2004~2006) [15] AKTIV-KAS Germany Traffic light violation (2006~2010) Right of way [16] Left turn; Right turn CVIS Europe Traffic light speed advice (2006~2010) Obstructed view [17] Permission denial Defective traffic signs Red light violation Emergency vehicle warning Intersafe2 Europe Left turn assistance (2008~2011) Right of way [2] Red light violation Right turn assistance Right turn for trucks

Required technologies Intervention DGPS unit, GPS, DAS, lasers canner, warning display Roadside sensors, Onboard units, GPS, digital map Roadside sensors, Onboard units, GPS, digital map DSRC, LDM15, Roadside sensors, onboard sensors

RSU, map, Stereo SRR; device

Active hazard braking system Active hazard braking system

GPS, ADAC Active hazLaser scanner; ard braking camera; Lidar; system Communication

However, little attention has been paid to priority and information conflicts. An information conflict occurs when information is received from various sources, such as a VMS, traffic lights, signals, and other vehicles. A priority conflict occurs when the same information is delivered to several vehicles at the same time. A mediator is needed to prevent or resolve these types of conflicts, and therefore this paper proposes an agent-based intersection system as a possible candidate.

3

Agent-based Intersection Safety Support System

The agent-based intersection support system is proposed to solve priority and information conflicts for improved safety. The system consists of an intersection agent, an infrastructure controller (agent), and vehicle agents, as shown in Fig 1. RSUs (roadside units) such as sensors and detectors notify the infrastructure controller that a vehicle has entered the intersection. The intersection agent takes the status information from the infrastructure controller and the vehicle agent, evaluates the information, and provides a warning when the risk of an accident occurs.

3

316

Fig. 1. Intersection context information model

An agent is a type of software with autonomy, sociality, learning, proactivity, and mobility characteristics [13]. Each intersection has an intersection agent and an infrastructure controller. The intersection agent supervises an intersection as a stationary agent, while the infrastructure controller manages the infrastructure, such as traffic signals, roadside sensors, and detectors. Each vehicle has a vehicle agent that, as a mobile agent, controls the vehicle and collects and identifies driver information. The proposed system has the following characteristics:       

Access to targeted service via unicast ID assignment considering privacy issues Identification of a driver’s tendencies Prediction of a driver’s decision during a risky situation Monitoring a driver’s action after receiving a warning Reevaluation of the accident risk Commands for collision avoidance VehicleAgent

IntersectionAgent

Collect data from vehicle analyze the driver tendencies Assign ID status report warning message predict the driver's action

InfrastructureController

Deliver the intersection infomation Deliver the detection information

Collect Data from infrastructure Detect a vehicle

Create ID Evaluate the risk of accidents 1 prepare to risk situation

update vehicle status Update the driver behavior information command messages for CV

vehicle status Evaluate the risk accidents 2 command messages for CV

action for CV

action for CA

Fig. 2. Sequence diagram for agent interaction model

4

317

Fig. 2 shows the agent-interaction process in detail. The vehicle agent collects information regarding the vehicular status and driver behavior, such as driving maneuvers, which is retrieved from the in-vehicle sensors. In addition, the agent identifies the driver’s tendencies based on the driver behavior information and predicts a driver’s action during a hazardous situation. The infrastructure controller collects data from the intersection infrastructure, such as traffic volume, speed, and signal intervals. It detects approaching vehicles by utilizing RSUs, and delivers the collected data to the intersection agent. The intersection agent monitors the situation at the intersection and evaluates the risk of an accident based on the information from the vehicle agent and the infrastructure controller. It warns the vehicle and orders the proper action for a collision avoidance based on the system policies. Point-to-point message delivery can make certain that only vehicles in dangerous situation will be warned.

4

Collision Avoidance in the Dilemma Zone

This section introduces the ‘dilemma zone’ and ‘type of collision’. In addition, it proposes an algorithm for collision avoidance in the dilemma zone. 4.1

Two types of collisions in the dilemma zone

A signalized intersection is controlled by a traffic signal control system, which consists of red, green, and yellow phases. A vehicle approaching a signalized intersection should decide whether to stop or proceed when the signal turns yellow. The dilemma zone is defined as the area approaching the stop line that is too close for a car to safely stop, and yet too far to pass fully through the intersection at a legal speed during a yellow light [9, 10]. If a driver decides to enter the intersection, the vehicle should pass clearly through distance ( ) at the permitted speed during the allowed time. The stopping distance ( is defined as the distance within which the vehicle can stop safely before the stop line if the driver chooses not to enter the intersection [9, 10]. Eq. (1) ( is the sum of the driving distance during the driver’s reaction time and the braking distance while stopping.

Fig. 3. Stopping distance: reaction distance + braking distance

,

(1)

where is the initial approaching speed of the vehicle, a is the vehicle’s braking deceleration, and is the driver’s braking response time. 5

318

(

Fig. 4. Clearing distance:

(2)

where is the remaining amount of yellow signal time, w is the width of the intersection, and L is the length of the vehicle.

Fig. 5. Dilemma zone (DZ): Stopping distance – Clearing distance

If , a dilemma zone exists, which is an area from which a vehicle can neither stop safely at the light nor completely pass through the intersection during the time allowed [18]. Any driver decision can lead to an accident or near-accident situation. Driver decisions can differ based on the driver’s perception of the situation, and the greatest variation in decisions occurs within the dilemma zone [10]. A driver may decide to stop or pass through an intersection while in the dilemma zone, and hence, two types of crashes can occur, a rear-end collision or a right-angle collision [12]. Such crashes may result from a sudden stop, accelerating during a yellow light, or passing through the intersection at a red light. A sudden stop may result in a rear-end or right-angle collision. A right-angle collision may also occur when the driver fails to clear an intersection while trying to pass through the intersection during a yellow light.

Fig. 6. Two types of collisions in dilemma zone

6

319

4.2

Algorithm for collision avoidance in the dilemma zone Collision Avoidance Algorithm in Dilemma Zone Step start

dstop – dgo>0 Warn the affected vehicles

Driver＇s decision?

Dilemma zone?

Collect data

Predict other risk of accidents

No

Yes Yes

Vt+k-Vt ≥ 0

Emergency warning to following vehicles

No

Rear-end collision?

V＇t+k-V＇t ≥ 0 Yes(ignorance warning) Yes

Dremain-dstop ≥ 0

Activate Brake system of the following vehicle No

Right-angle collision?

Stop in stop line

Command No signal control

dremain - dgo

Command Pass message

≥0

Yes Clear intersection

stop

Fig. 7. Collision avoidance algorithm

Fig. 7 shows the details of the algorithm used for collision avoidance. First, the agents collect information. A vehicle agent collects data, such as the speed, direction, and location of a vehicle. In addition, it collects information about the driver’s behavior. The driver’s tendencies are analyzed based on this information. An intersection agent collects static information, such as the intersection structure, signals, and fixed signal intervals. An infrastructure controller collects dynamic information, such as the entering vehicles, traffic flow, and traffic volume. If the infrastructure controller detects an approaching vehicle, it notifies the intersection agent. Second, the system warns the vehicles in the dilemma zone in advance. The intersection agent evaluates whether an approaching vehicle will enter the dilemma zone. If , the vehicle will enter the dilemma zone. The agent 7

320

then sends a warning to the vehicle agent, which prepares for a potentially dangerous situation based on the driver’s tendencies. Next, the system monitors the driver’s action. The system monitors the speed change of the vehicle after the warning is received. If the velocity of the vehicle after the driver’s response ( ) does not decrease, it indicates that the driver has decided to go through the intersection. If the velocity decreases, the driver’s decision is to stop. is the velocity of the vehicle before it has been warned. If , the driver has decided to cross the intersection. The system then reevaluates the risk of an accident based on the driver’s decision.  If the driver decides to stop, there is the risk of a rear-end or right-angle collision. If , the system warns the following vehicles to stop. ─ Rear-end collision If the driver decides to stop, but the driver of the following vehicle decides to continue, a rear-end collision may occur. Therefore, the system should monitor the velocity of the following vehicle ( . is the speed of the following vehicle after receiving a warning, while is the following vehicle’s velocity before a warning is received. If > 0, the system commands an action, such as activating the brake system of the following vehicle, depending on the system’s policy for avoiding rear-end collisions. ─ Right-angle collision If is not sufficient, but the driver decides to stop, the vehicle may miss the stop line, creating the risk of a right-angle collision. If the remaining distance ( ) to the stop line is greater than the stopping distance, the vehicle can stop safely. However, if , the system warns the affected vehicle, and controls the subsystems based on a pre-determined policy.  If the driver decides to go through the intersection, the vehicle is at risk for a rightangle collision. To cross through the intersection clearly, the remaining distance ( ) to the clear line should be shorter than the clearing distance during the allowed amount of time. If , there is a risk of a right-angle collision. ─ Right-angle collision If 0, the system can make by controlling a subsystem such as the traffic signal based on the policy of the intersection system. Various actions can be taken for collision avoidance depending on the policy, such as activating a braking system or adjusting the signal interval. For the purpose of this paper, however, it is not necessary to discuss such policies in detail. 8

321

5

Demonstration

This study has developed a prototype system by utilizing the JADE (Java Agent DEvelopment) Framework. This section describes a simple interaction between an intersection agent and a vehicle agent used in the system.

Fig. 8. Demonstration of collisions in dilemma zone

The left image in Fig. 8 shows vehicles in the dilemma zone. The first vehicle has missed the stop line, which may result in a right-angle collision. The center image shows the occurrence of a rear-end collision. Finally, the right image is a depiction of a right-angle collision. Fig.9 shows example messages between agents to avoid the collisions depicted in Fig. 8.

Fig. 9. Vehicle agent collision warning test

6

Conclusion

This paper proposed an agent-based intersection system for avoiding accidents and improving intersection safety. In addition, it suggested the use of an algorithm for collision avoidance in the dilemma zone. The agent-based technology is applied to vehicles and intersections to collect and deliver information to drivers. The system analyzes the risk of an accident, and takes appropriate action depending on the situation. The proposed system solves priority and information conflicts though interaction

9

322

and cooperation among agents. It also provides a targeted service for each vehicle during a dangerous situation.

References 1. Traffic Accident Analysis System, http://taas.koroad.or.kr 2. INTERSAFE-2: Cooperative Intersection Safety, http://www.intersafe-2.eu 3. Chen, B., Cheng, H.: A Review of the Applications of Agent Technology in Traffic and Transportation Systems. In: IEEE Transactions On Intelligent Transportation Systems, Vol. 11, No. 2, pp. 485 --497, IEEE (2010) 4. Chen, B., Cheng, H., Palen, J.: Integrating mobile agent technology with multi-agent systems for distributed traffic detection and management systems. In: Transportation Research Part C. 2009. pp.1--10. Elsevier (2008) 5. Delozier, G.: Ford's "Smart Intersection" Could Reduce Intersection Accident.(2011) 6. 경찰청.: 2011 교통사고통계, 범신사 (2011) [translation: Korean National Police Agency.: the 2011 traffic accident statistics (2011)] 7. Salim, F.D., Cai, L., Indrawan, M., Loke, S.W.: Road Intersections as Pervasive Computing Environments: towards a Multiagent Real-Time Collision Warning System, pp 621, In: 6th IEEE International Conference on Pervasive Computing and Communications, IEEE, (2008) 8. http://wiki.answers.com/Q/What_is_a_dilemma_zone 9. Puan, O.C., Ismail, C.R.: Dilemma Zone Conflicts at Isolated Intersections Controlled with Fixed–time and Vehicle Actuated Traffic Signal Systems. In: International Journal of Civil & Environmental Engineering IJCEE-IJENS Vol.10 No.03, pp. 19—24. (2010) 10. Zimmerman, K., Bonneson, J.A.: Intersection Safety at High-Speed Signalized Intersections: Number of Vehicles in Dilemma Zone as Potential Measure, In: Transportation Research Record: Journal of the Transportation Research Board, No. 1897, pp.126--133 TRB, National Research Council, Washington, D.C., (2004) 11. Urbanik, T., Koonce, P.: The Dilemma with Dilemma Zones. In: Proc., ITE District 6 Annual Meeting, (2007) 12. Tang, S.M.: A preliminary investigation on artificial transportation systems. In: J. Syst. Simul., vol. 17, no. 3, pp. 704–709, (2005) 13. Li, J., Tang, S., Wang, X., Duan, W., Wang, F.: Growing Artificial Transportation Systems: A Rule-Based Iterative Design Process. In: IEEE Transactions On Intelligent Transportation Systems, Vol. 12, No. 2, IEEE (2011) 14. Cooperative Intersection Collision Avoidance Systems (CICAS), www.its.dot.gov/cicas 15. Smartway 2007 (Global ETC and Japanese ITS Information), http://www.hido.or.jp/itsos/ 16. AKTIV, http://www.aktiv-online.org 17. CVIS, http://www.cvisproject.org 18. Mcshane, W.R., Roess, R.P., Prassas, E.S.: Traffic Engineering 2nd edition, pp.195, 442— 443, Prentice Hall (1988)

10

323

Sensor Data Network Using Android Mobile Devices Chris Jordan G. Aliac Intelligent Systems Laboratory, Cebu Institute of Technology, Cebu City, 6000 Philippines [email protected]

Abstract. We have an insatiable desire to make sense of the world around us. How do we best observe and record the details of time, nature, location, events, and our own personal experiences? Data collection has always been a major challenge in sensor networks and various techniques have been proposed to en- able efficient data collection. One such methodology is the use of mobile ele- ments within the existing infrastructure to enable data collection. Smartphone’s are becoming ever more popular as they become cheaper to buy and very user friendly and easy to write applications. A wide range of these smartphones’ are based on the Android operating system. In this paper, we present a system that deploys lightweight Sensor Networks easily based on widely available Android Devices. An Android application is used to communicate with the on-board sensors of the mobile device, collate the variety of sensor readings to then be displayed on the device and sent to a central server for data logging and analy- sis. To ensure data reliability and redundancy, back up logs are saved within the sensor devices and then later sent up the server or other live sensor device via SMS message.

1

Introduction

Research into semantic sensor networks has been focusing on the treatment and processing of data aggregated from large networks of sensors, often based on specialized equipments geographically distributed in large areas, [1] discusses a number of challenges related to Semantic Sensor Networks in such scenarios. The challenge we are particularly interested in relates to the ability for rapid development of applications that make use of Sensor Network. We especially look at applications in scenarios where it is needed to set -up networks of simple sensors quickly and easily (e.g., school projects, small experiments).In recent years, the Android platform became a de-facto standard for different types of mobile devices from several manufacturers. The computational power of these devices already allows client processing of small to medium volumes of semantic data [2]. According to Gartner Press Release for November 10, 2010 [3], Android is one of the most popular smartphone platforms at the moment, and the popularity is even rising. It is one of the most open and flexible platforms providing software developers easy access to phone hardware and rich software API. Android mobile devices not shows as a powerful computing and communication mobile device of choice, but it also comes with a

324

rich set of embedded sensors, such as an accelerometer, digital compass, gyroscope, GPS, microphone, and camera. Collectively, these sensors are enabling new applications across a wide variety of domains, such as healthcare [4], social networks [5], safety, environmental monitoring [6], and transportation [7, 8], and give rise to a new area of research called mobile phone sensing. Until recently mobile sensing research such as activity recognition, where people’s activity (e.g., walking, driving, sitting, talking) is classified and monitored, required specialized mobile devices (e.g., the Mobile Sensing Platform [MSP]) [9] to be fabricated [10]. Mobile sensing applications had to be manually downloaded, installed, and hand tuned for each device. User studies conducted to evaluate new mobile sensing applications and algorithms were small-scale because of the expense and complexity of doing experiments at scale. As a result the research, which was innovative, gained little momentum outside a small group of dedicated researchers. Although the potential of using mobile phones as a platform for sensing research has been discussed for a number of years now, in both industrial [11] and research communities [12,13], there has been little or no advancement in the field until recently.

2

System Overview

Fig. 1. Data Collection Overall Layout

A sensor network is a data centric network which comprises distributed sites of

325

information represented by sensor nodes contributing data sensed [14]. They are producer of valuable data that needs to be collected in a cost efficient manner. The cost includes cost of communication, power consumption and time involved in communication. Data collection can be classified into Data Collection using static nodes and Data Collection using mobile elements. Since our focus is on mobile device based data collection, the literature looks into the use of mobile elements for data collection and not into traditional based multi hop and clustering based techniques. The idea of using mobility within sensor network to collect data from sensor nodes has been discussed by a number of researchers [15, 16, 17, 18, 19, 20, 21, 22]. The use of mobility to improve performance in AdHoc networks has been proposed in ZebraNet project and Manatee project [3]. These mobile elements can be introduced or existing mobility in the environment can be used to achieve mobility. The use of mobility can be classified broadly as Random mobility, Predictable mobility and Controlled mobility. Kansal et al. [4, 12] presents a work on using controlled mobility for data col- lection in sensor networks. Shah et al. [3] and Sushant et al [12] present Data Mules that makes use of random mobility for data collection. Chakrabarthi et al. [11] proposes a novel way of using predictable mobility to collect data from sensor networks that are distributed over an area.

2.1

Accelerometer Sensor

As stated in the above literature, Android devices contain sereval on-board sensors which can be easily acquired by writing an Android application. In this paper, the accelerometer sensor, which is always found on Android phones, will be demonstrated as a potential data on which to acquire. The devices’ accelerometer consists of 3 axes (X, Y, Z). Forces are applied to the sensor which yields results in m/s2. Normally gravity acts upon the sensor and depending on the phones position, one of the phones axes will register 9.8 m/s2 at stationary nonmoving state.

Ad = - ∑Fs / mass Fig. 2. 3 Axis accelerometer forces

326

By measuring accelerometer data, can yield a lot of potential applications such as vibration sensing on a specific rigid structure, measuring road conditions traveled by car and measuring seismic activity to name a few.

2.2

Android SDK

Android is a software stack for mobile devices that includes an operating system, middleware and key applications. T he Android SDK provides the tools and APIs necessary to begin developing applications on the Android platform using the Java programming language [23].

Fig. 3. Adroid OS Stack

By providing an open development platform, Android offers developers the ability to build extremely rich and innovative applications. Developers are free to take advantage of the device hardware, access location information, run background services, set alarms, add notifications to the status bar, and much, much more. Developers have full access to the same framework APIs used by the core applications. The application architecture is designed to simplify the reuse of components; any application can publish its capabilities and any other application may then make use of those capabilities (subject to security constraints enforced by the framework). This same mechanism allows components to be replaced by the user. Underlying all applications is a set of services and systems, including:

327

A rich and extensible set of Views that can be used to build an application, including lists, grids, text boxes, buttons, and even an embeddable web browser Content Providers that enable applications to access data from other applications (such as Contacts), or to share their own data A Resource Manager, providing access to non-code resources such as localized strings, graphics, and layout files A Notification Manager that enables all applications to display custom alerts in the status bar. An Activity Manager that manages the lifecycle of applications and provides a common navigation backstack Android includes a set of C/C++ libraries used by various components of the Android system. These capabilities are exposed to developers through the Android application framework. Some of the core libraries are listed below: System C library - a BSD-derived implementation of the standard C system library (libc), tuned for embedded Linux-based devices Media Libraries - based on PacketVideo's OpenCORE; the libraries support playback and recording of many popular audio and video formats, as well as static image files, including MPEG4, H.264, MP3, AAC, AMR, JPG, and PNG Surface Manager - manages access to the display subsystem and seamlessly composites 2D and 3D graphic layers from multiple applications LibWebCore - a modern web browser engine which powers both the Android browser and an embeddable web view SGL - the underlying 2D graphics engine 3D libraries - an implementation based on OpenGL ES 1.0 APIs; the libraries use either hardware 3D acceleration (where available) or the included, highly optimized 3D software rasterizer FreeType - bitmap and vector font rendering SQLite - a powerful and lightweight relational database engine avail- able to all applications

2.3

Data Acquisition Infrastructure

In this Sensor network, a server runs as the data acquisition for the different sensors. The System is connected via WIFI connection in which the server opens up a TCP socket port. The Android mobile sensors will act as clients connected to the TCP port. Given the simplicity of the way devices connect in such a way that TCP runs on any platform, servers may appear from different device types of different platforms. Server devices may appear as a desktop computer, a laptop or even a tablet PC. Platforms run from Windows based or even Linux or MAC Systems. In this paper, Windows platyform is run using .Net C# as the programming language of choice.

328

3

Sensor Data Network Potentials

Although the system shows limitations to the number of sensors connected to the system. The architecture can be expanded to form a network of networks. TCP network servers can act as a large sensor array and can communicate to other TCP network servers. It can also be altered to become a Semantic Sensor Data Network by altering ports to Internet (port 80) and establishing Web servers for the TCP severs to send data or even better, make TCP server to WEB servers and alter the Android sensor apps to connect via Internet Web.

3.1 Reliability Data from sensors must be sent to servers by means of a standard TCP socket. But there will be instances wherein TCP servers will be offline. At this condition, to ensure data will be sent to the server, Data will also be backed-up on the Android mobile sensors. Upon a substantial amount of time will be given for the server to be back online, when the server remains offline, data will be sent to other Android mobile sensors via SMS. A redundant SMS server can also be established online for the mobile sensors to tramsit data. In a worst-case failed network scenario.

4

System Results

Two Android devices were installed with a written application that acquires data from its accelerometer and connects to a server program funning on the WIFI network. Data is tramsitted in the form of raw float data from the 3 axis. Data received form the server is displayed onto the UI and options for saving the data is created. A graph is also displayed from the given realtime data in separate axis values.

4.1 Android Application program The android application program was written using the Eclipse IDE. Since accelerometer is not possible in IDE Android emulator, an actual Android phone was connected and program was tested via USB connections an run on Android debug mode ADB.

329

Fig. 4. Android Application

4.2

Server Program

The server program was written using Visual C# express edition. Dot Net libraries provide a powerful socket connection library via TCP Listeners. The Server runs on port address 8000 and can easily be ported to port 80 internet connection.

330

Fig. 5. TCP Server Progam

4.3

Data Analysis

Since Data is passed to the server as 3 xis data in the form below: sensor1: X:-1.493966 Y:-0.306454 Z: 9.998084 Data can be analyzed by the sensor sending the data and the current real time data by the 3 axis X, Y and Z. The data shown above tells us that the phones 2 position is lying flat on its back by the Z data at 9.9 m/s . It would lie face first if 2 the data reflect -9.9 m/s . Based on initial trials of data logged and saved, sensor data is very sensitive that a simulated movement test was conducted. Data was 2 taken and absolute offset from 9.9 m/s was calculated .Results are shown below. Table 1. : Simulated movement

test

331

Although the table shown on the table does not show conclusive data due to the fact that no basis from actual data was used to conclude remarks. This table shows that raw data results can be easily analyzed for a specific study.

5

Conclusion

Based on the system results, this system can of potential use to different applications for different disciplines. The network system can be easily converted form a basic data sensor network where basic sockets are used on TCP, to a full Semantic sensor web that could gather a large array of data.

References 1. Corcho, O., Garcia-Castro, R.: Five challenges for the semantic sensor web. Semantic Web Journal 1(1) (2010) d'Aquin, M., Nikolov, A., Motta, E.: How much semantic data on small devices? In: EKAW. (2010) 3. http://www.gartner.com/it/page.jsp?id=1466313 4. S. Consolvo et al., “Activity Sensing in the Wild: A Field Trial of Ubifit Garden,” Proc. 26th Annual ACM SI GCHI Conf. Human Factors Comp. Sys., 2008, pp. 1797–1806. 5. E. Miluzzo et al., “Sensing meets Mobile Social Networks: The Design, Implementation, and Evaluation of the CenceMe Application,” Proc. 6th ACM SenSys, 2008, pp. 337–50. 6. M. Mun et al., “Peir, the Personal Environmental Impact Report, as a Platform for Participatory Sensing Systems Research,” Proc. 7th ACM MobiSys, 2009, pp. 55–68. 7. A. Thiagarajan et al., “VTrack: Accurate, Energy-Aware Traffic Delay Estimation Using Mobile Phones,” Proc.7th ACM SenSys, Berkeley, CA, Nov. 2009. 8. UC Berkeley/Nokia/NAVTEQ, “Mobile Millennium”;http://traffic.berkeley.edu/ 9. T. Choudhury et al., “The Mobile Sensing Platform: An Embedded System for Activity Recognition,” I EEE Pervasive Comp., vol. 7, no. 2, 2008, pp. 32–41. 10. T. Starner, Wearable Computing and Contextual Awareness, Ph.D. thesis, MIT Media Lab., Apr. 30, 1999. 11. Nokia, “Workshop on Large-Scale Sensor Networks and Applications,” Kuusamo, Finland, Feb. 3–6, 2005. 12. A. Schmidt et al., “Advanced Interaction in Context, ”Proc. 1st Int’l. Symp. HandHeld Ubiquitous Comp., 1999, pp. 89–101. 13. N. Eagle and A. Pentland, “Reality Mining: Sensing Complex Social Systems,” Personal Ubiquitous Comp., vol. 10, no. 4, 2006, pp. 255–268. 14. Chien-Chung Shen, Chavalit Srisathapornphat, and Chaiporn Jaikaeo. Sensor inform ation networking architecture and applications. IEEE Personal Communications, 2001, Page(s): 52-59 15. R. Shah, S. Roy, S. Jain, W. Brunette, "Data MULEs: Modeling a Three-tier Architecture for Sparse Sensor Networks", IEEE SNPA Workshop, May 2003 16. K. Aman, A. S. Arun, D. J. David, B. S. Mani, and E. Deborah, "Intelligent fluid infrastructure for embedded networks," in Proceedings of the 2nd international conference on Mobile systems, applications, and services. Boston, MA, USA: ACM Press, 2004. 2.

332

17. David Jea, Arun Somasundara, Mani Srivastava, “Multiple Controlled Mobile Elements (Data Mules) for Data Collection in Sensor Networks”, International Conference on Distributed Computing in Sensor Systems, 2005 18. LaMarca, A., Brunette, W., Koizumi, D., Lease, M., Sigurdsson, S. B., Sikorski, K., Fox, D., and Borriello, G. Making Sensor Networks Practical with Robots. In Proceedings of the First international Conference on Pervasive Computing, 2002, Page(s): 152-166. 19. Shashidhar R. Gandham, M. Dawande, Ravi Prakash and S. Venkatesan, Energy-Efficient Schemes for Wireless Sensor Networks with Multiple Mobile Base Stations. Proceedings of IEEE Globecom, December 2003 20. Vincze, Z., Vass, D., Vida, R., Vidács, A., and Telcs, A. Adaptive sink mobility in eventdriven multi-hop wireless sensor networks. In Proceedings of the First international Conference on integrated internet Ad Hoc and Sensor Networks, 2005 21. Tong L, Q. Zhao, and S. Adireddy, "Sensor networks with mobile agents," Proc. Of IEEE MILCOM'03, 2003 22. Ren, B., Ma, J., and Chen, C. The hybrid mobile wireless sensor networks for data gathering. In Proceeding of the international Conference on Communications and Mobile Computing, 2006 23. http://developer.android.com/guide/basics/what-is-android.html

333

Ontology-based RFID code conversion for ubiquitous applications Young-Sik Noh, Yung-Cheol Byun, SangJoon Lee Dept. of Computer Engineering, Jeju National University, 66 Jejudaehakno, Jeju-Si, Jeju-Do, 690-756, Korea {solyo,ycb,sjlee}@jejunu.ac.kr

Abstract. RFID is one of the important techniques to implement ubiquitous applications and services. To process various kinds of RFID readers and sensors, we need an RFID middleware system. ALE, which is the name of RFID middleware proposed by EPCglobal, plays an important role as the middleware system. The system processes EPC code data and provides it to ubiquitous services. In this paper, we extended the system to effectively process with some number of newly emerged EPC code data. Based on the experimental results, we found that the system could operate properly to process the code data. Keywords: Ontology, Ubiquitous services, Semantic, EPC code

1

Introduction

To implement various kinds of ubiquitous services with good quality, we need a number of core technologies, and RFID is one of the important techniques [1][2]. RFID middleware exists in the center of the technology from the view point of software systems. EPCglobal, a well known organization leading international standards, proposed an international standard of 900MHz, EPC Class 1 Gen2 which was selected as ISO 18000-6C, as well. The organization also proposed EPC Network Architecture, which has been adapted as a standard by a number of companies doing their business with RFID middleware systems. In the architecture, ALE which is the name of middleware proposed by the organization plays an important role as an RFID middleware with some important functions including filtering, collecting, reporting, and etc. The middleware system processes EPC code data and provides it to ubiquitous services. That is, ALE collects EPC data from a reader device and processes it properly. The system internally converts the EPC code into URN type and performs filtering, grouping, and other operations to provide the result to ubiquitous services. Because there are many kinds of EPC codes, we proposed the system to process [2] effectively. In this paper, we extended the system to effectively process with some number of newly emerged EPC code data.

334

2

Lecture Notes in Computer Science: Authors’ Instructions

This paper is organized as follows; In Section 2, we invest some related works and systems. In Section 3, we present the method to extend our EPC code processing system. We implement and verify the system in Section 4, and conclude the research in Section 5.

2

Related Works

We can find some representative RFID middleware systems [2] including Sensor Edge Server by Oracle [3], Java RFID system by Sun Microsystems [5], TagsWare by CapTech [4], and XML-based sensor system [6]. Sensor Edge Server is a sensor based service integration platform with data acquisition, event processing, and data dispatching as main functions. TagsWare consists of links to send RFID tag data to applications, drivers as a standard interface between RFID reader devices and applications, and base components which enable applications to use the links and drivers. After gathering and smoothing raw data from reader devices, the system sends the result to applications by using link components connected to each other in a chain form. Java RFID system is an event manager to process sensor data stream with reader adapter, filter, logger, enterprise gateway. By using a query issued by a user, the system extracts EPC tag data satisfying specific mask conditions. XML-Based sensor system manages scattered diverse sensors, collects data from sensors, processes collected sensor data, and automatically delivers the processed results to related ubiquitous services.

3

Code conversion based on Ontology

First, we designed and extended our ontology architecture. To represent all kinds of code by using Ontology, we defined Code as a parent class [2]. Classes such as ISO, Bar-code, and EPC are defined as a sub-class of the Code class in addition to related attributes, which represents ISO international standard code, EPC code by EPCglobal, and bar-code being currently used. To convert many kinds of EPC code types effectively, the code conversion rule for each code is needed. So, some rules for EPC code conversion were added in the next step. The rules present the structural information according to each ECP code type, which is used to convert a collected code. From now on, the system can generate the URN code by using the rules defined with ontology. EPC code collected from a reader is converted first into binary code, and the value of Header field is extracted. By using the value of Header field as a key, the system finds a conversion rule instance from the rule ontology for code conversion. From the instance, the name of EPC code, size, URN header in AdditionalInfo is extracted and the location information for the entire field is recognized as well.

335

Lecture Notes in Computer Science: Authors’ Instructions

4

3

Experimental results

We implemented the code conversion rules with ontology to process RFID codes including newly defined codes and tested if the system operates well. We tested the system by using the experimental dataset composed of the existing 12 types of Gen1/Gen2 codes and five types of newly emerged codes. By using the experimental dataset, we tested the main functions of the system including analyzing the header of EPC tag and conversion with the rules defined with the ontology database. For a code of SGLN-64 such as 8000000070020000, the system correctly found the header value and retrieved the structural information from ontology by using the header value as a key. The binary code data was segmented and converted into URN code format such as urn:epc:tag:sgln-64:0.0.37.60002. For a code of SGTIN-96, the system found the header bit string such as 00110110, and the code was converted into the URN code of SGTIN-96 such as urn:epc:tag:sgtin198:0.0.18432.0.ER3. Experimental results showed the conversion of 17 types of EPC codes into URN codes by using the proposed method. A total number of 170 tag dataset where 10 instances of tag data for each type of EPC data were used. From the results, we verified that all of the codes could be converted successfully.

5

Conclusion

In this paper, we extended our conversion method of EPC tag data into URN code to process a new emerged RFID codes. With the method, ALE-compliant middleware can process new types of EPC data collected from RFID devices. Especially, we extended the existing ontology database defining the rule and knowledge to convert RFID code. To evaluate our approach, we tested its performance by using 170 instances of tag data included in the existing twelve different types of EPC code and newly emerged five types. Based on the experimental results, we found that the system operates properly to process RFID codes.

Acknowledgements This research was supported by the MKE (The Ministry of Knowledge Economy), through the project of Region technical renovation, Republic of Korea.

References 1. Young-Sik Noh, Yung-Cheol Byun: Intelligent Home Services based on Contextawareness. Journal of Maritime Information and Communication, vol. 11, No. 4, pp. 678–686 (2007)

336

4

Lecture Notes in Computer Science: Authors’ Instructions

2. Yung-Cheol Byun, Young-Sik Noh, Sang-Young Byun, Chul-Ung Kang, Jong-Hwan Lim: Automatic Code Conversion based on Ontology for Extensible RFID Middleware. In: 3rd International Conference on Ubiquitous Information Technologies, Vietnam (2008) 3. Enterprise Information Architecture for RFID and Sensor-Based Services, Oracle White Paper (2006) 4. TagsWare: Agile RFID Solutions, http://www.tagsware.com 5. A. Gupta, M. Srivastava: Developing Auto-ID Solutions using Sun Java System RFID Software, http://java.sun.com/developer/technicalArticles/Ecommerce/rfid/sjsrfid/RFID.html (2008) 6. H. S. Lee, H. H. Choi, B. S. Kim, M. C. Lee, J. H. Park, M. Y. Lee, M. J. Kim, S. I. Jin: UbiCore : An Effective XML - based RFID Middleware System. Journal of Korea Information Science Society, vol. 33, pp. 578–589 (2006)

337

Location Recognition of Wireless Sensor Nodes Based on Environmental Data Classification Ae-cheoun Eun and Young-guk Ha* Department of Computer Science & Engineering, Konkuk University, Seoul, Korea [email protected] and [email protected] *Corresponding Author

Abstract. In ubiquitous environments, sensor localization plays a critical role in various location-based systems. Commonly, sensor location is measured with Global Positioning System (GPS). Because GPS is not available in indoor environments, Received Signal Strength (RSS) of wireless sensors is used for indoor localization in place of GPS. However, there is a problem in measuring sensor location using RSS because of radio signal interference caused by various indoor obstacles. Thus, we used measured environmental data and a classification method to recognize location where the sensor node belongs. In this paper, we used a wireless sensor node to collect environmental data about temperature, humidity, sound, and light of each location. Then, we extracted some features from the collected data and trained the location data classifier to recognize location of the wireless sensor node. Keywords: Environmental Sensor, Location Data Classification, Localization.

1

Introduction

Localization for wireless sensor network (WSN) becomes an essential functionality because location-aware service is an important application of ubiquitous computing. Basically, localization of WSN is done by measuring Received Signal Strength (RSS) of wireless links between a target node and multiple reference nodes based on the theory that the signal strength decreases with distance. The measured RSS data are used to determine location of the target node with the triangulation [1], centroid method [2], or fingerprinting [3]. However, such a scheme has some limitations in application to indoor environments because of signal reflection, loss and distortion with various indoor obstacles. In this paper, we propose a novel indoor location recognition scheme for Zigbee sensor nodes using classification of environmental data, such as temperature, humidity, sound, and light data, measured by the target node. To classify those environmental data for each location, we used a k-Nearest Neighbor (kNN) classifier. Then we performed localization experiments in the actual test environment to validate the proposed scheme.

338

2

Proposed Localization Scheme

Fig. 1 shows our system architecture and data flow. The Location Data Collection Module (LDCM) periodically provides environmental data of each space by sensing and collecting. Collected environmental data of each space is used for User Location Recognition Module (ULRM) training. Also LDCM can sense current environmental data at user location. Finally sensed data will be used to recognize user’s location as test data.

Fig. 1. Architecture of the proposed location recognition system.

The location data feature extraction module (LFEM) provides feature extraction function from environmental data of user location. Extracted feature can be used in user location recognition at use ULRM. In addition LFEM provides different way of extraction from each feature. ULRM is a set of trained data for recognizing location. Also ULRM shows result of recognition based on extracted real-time data from LFEM.

Fig. 2. Detailed architecture of User Location Recognition Module.

Fig. 2 shows detailed architecture of ULRM. ULRM uses the space recognition features generated from LFEM for training. Location data classifier is used to classify the current user’s location features. To this end, location data classifier should be training based on the set of environmental data. ULRM input through the user location recognizer classification is based on the received environmental data to provide a result.

339

3

Implementation and Experiment

For the implementation of the proposed location recognition system, we used some software tools and Zigbee sensor nodes listed in table 1. Table 1. Implementation of the proposed location recognition system. Operating system Programming language Software tools Hardware

- Location recognition system: Windows Vista - Wireless sensor node: TinyOS - Location recognition system: Java - Wireless sensor node: nesC - Location data feature extraction: MATLAB - Location data classification: Weka - Location recognition system: Intel 2.0Hz PC - Wireless Sensor node: Hmote2420

In our experiments, the environmental data is collected at the campus of Konkuk University. Fig. 3 shows the places where the environmental data are collected. Our experimental method is as follows. At first, we collect 100 datasets from each location with environmental sensors. And the collected datasets are input to feature extraction method of MATLAB and the output of the feature extraction is converted into a dataset for kNN data classifier of Weka. Next, we collect 10 more datasets from the same locations and use the datasets to test the location data classifier.

Fig. 3. Detailed architecture of User Location Recognition Module.

Table 2 shows the confusion matrix for the offline test results. The 3-NN classification method with a 20-fold cross validation was used for this experiment. As shown in the matrix, the overall localization accuracy is about 95.3%. This table shows that the most well recognized places are laboratory and cafeteria. Table 2. Confusion matrix of the offline localization experimental results. Classified Test Data Lobby Laboratory Toilet Cafeteria Bank Bookstore

Lobby

Laboratory

Toilet

Cafeteria

Bank

Bookstore

91 0 1 0 4 0

0 99 0 0 0 3

1 0 94 0 2 0

0 0 0 99 2 0

8 0 5 1 92 0

0 1 0 0 0 97

340

Table 3 shows the real-time location recognition rate. Real-time location recognition in experiment showed that average is 82.2%. And in this experiment, the highest recognition rate was shown in the toilets. The bookstore, on the other hand, shows the lowest recognition rate. The classifier confused bookstore with the lobby. This problem occurred because both locations can have similar light and temperature conditions. However, in the case of the toilet, because of the high humidity, the result showed a high probability. Table 3. Real-time localization experimental results. Location Laboratory Lobby Toilet Cafeteria Bank Bookstore Average

4

Recognition rate 76.7% 83.3% 100% 86.7% 93.3% 53.3% 82.2%

Conclusion

In this paper, we proposed a novel location recognition method for wireless sensor nodes using classification of environmental data feature with the kNN localization data classifier. And we performed localization experiments in our actual test environment using the proposed location recognition method. Finally, our future research will be focused on combining the proposed location recognition method and other localization methods such as RSS pattern recognition. Acknowledgement.This research was supported by Basic Science Research Program through the National Research Foundation of Korea(NRF) funded by the Ministry of Education, Science and Technology(grant number: 2012006817).

References 1. Xia Zhenjie and Chen Changjia, “A Localization Scheme with Mobile Beacon for Wireless Sensor Networks,” The 6th Intermational Conference on ITS & Telecommunications, 2006. 2. Chin-Heng Lim, Yahong Wan, Boon-Poh Ng, Chong-Meng Samson See, “A Real-Time Indoor WiFi Localization System Utilizing Smart Antennas,” IEEE Transactions on Consumer Electronics, Vol. 53, No. 2, May. 2007. 3. Ae-cheoun Eun and Young-guk Ha, “An Indoor Localization Scheme for Zigbee Sensor Networks Using LQI Fingerprinting,” International Conference on Advanced Information Technology and Sensor Application, 2012.

341