Fuzzy Fusion in Multimodal Biometric Systems - Semantic Scholar

Fuzzy Fusion in Multimodal Biometric Systems V. Conti1, G. Milici1, P. Ribino1, F. Sorbello1 and S. Vitabile2 1

Dipartimento di Ingegneria Informatica Università degli Studi di Palermo Viale delle Scienze, Ed. 6, 90128 Palermo, ITALY {conti, milici, sorbello}@unipa.it 2 Dipartimento di Biotecnologie Mediche e Medicina Legale, Università degli Studi di Palermo Via del Vespro, 90127 Palermo, ITALY [email protected]

Abstract. Multimodal authentication systems represent an emerging trend for information security. These systems could replace conventional mono-modal biometric methods using two or more features for robust biometric authentication tasks. They employ unique combinations of measurable physical characteristics: fingerprint, facial features, iris of the eye, voice print, hand geometry, vein patterns, and so on. Since these traits are hardly imitable by other persons, the aim of these multibiometric systems is to achieve a high reliability to determine or verify person's identity. In this paper a multimodal biometric system using two different fingerprints is proposed. The matching module integrates fuzzy logic methods for matching score fusion. Experimental trials using both decision level fusion and matching score level fusion were performed. Experimental results show an improvement of 6.7% using the matching score level fusion rather then a mono-modal authentication system. Keywords: Fuzzy fusion, fingerprint, afas, matching score level fusion.

1

Introduction

A biometric system regards a pattern recognition system acquiring biometric data from a user, extracting features from it, and comparing extracted features with stored template sets. Traditionally biometric systems, operating on a single biometric feature, have many limitations [14]: • troubles on data sensor: sensor captured data are often affected from “noise” due to the environmental conditions (insufficient light, powder etc) or to user physiological conditions (cold, cut fingers etc); • distinctiveness ability: all biometric features have not the same distinctiveness (for example, hand geometry biometric systems are less selective than those based on fingerprints);

2

V. Conti at al.

•

not universality: every biometric feature is universal, but some people not owned them for some physical invalidity (lack of a finger or the voice). Multimodal biometric systems [6][7][8] are a new approach to overcome these problems. These systems promise significant improvements over biometric system using a single characteristic, in terms of higher accuracy and most resistance to spoofing. In literature different approaches for multimodal biometric systems have been proposed. These systems are based on different biometric features or introduce different fusion algorithm of these features. In [8] Jain et al. have proposed a multimodal biometric system which integrates face-fingerprints-voice recognition modules to realize personal identification system. The authors have used a statistical decision theory based on Neyman-Pearson rule to classify a given observation. In [9] Lau et al. have presented a multimodal biometric system combining facevoice-fingerprints. The authors used a fuzzy logic based approach, in order to consider the effect of external conditions on the system. With more details they have implemented fuzzy logic module to calculate the weights for each recognition subsystem to realize the weight sum rule. In [10] Zewail et al. have presented an identity verification system based on soft and hard biometrics. The authors proposed a fingerprint-iris based system integrating extra soft biometric information, as the colour of the iris, to enhance its performance. In this paper a multimodal biometric system, combining two fingerprint authentication systems, is proposed. With the proposed approach some biometric monomodal authentication systems limitations has been reduced. The used fusion techniques are based on fuzzy logic considering fingerprints image quality as in particular, two fusion methods have been tested: the first one implements a decision level fusion, the second one implements a matching score fusion. The proposed approach uses the images quality as goodness index, as the authors in [9], but nothing weight is obtained from fuzzy rules to characterize the system phases. The developed approach can be adopted with general multimodal authentication systems involving different biometric features. The paper is organized as follows. Section II deals with taxonomy of multimodal systems; section III introduces the proposed system; in the section IV the experimental results are reported; and finally in section V some conclusions are described.

2

Taxonomy of Multimodal Systems

The multimodal biometric systems can be classified considering the biometrics source number and the biometrics samples number [7][11]. Four different models of multimodal systems can be distinguished: • the MSSS (MultiSample-SingleSource) model: many samples of one single biometric feature are used to recognize a user; • the SSMS (SingleSample-MultiSource) model: only one sample of many biometric features is used;

Fuzzy Fusion in Multimodal Biometric Systems

3

•

the MSMS (MultiSample-MultiSource) model: many sensors for every biometric feature are used; • the multi-algorithms model: different matching algorithms are used in the recognition of one single sample of one biometric feature. The proposed multimodal approach implements the SSMS model.

3

The Proposed Multimodal Biometric System

3.1

Standard Fingerprint Authentication

In general, an automatic fingerprint authentication system [1][2][12], commonly called AFAS (Automatic Fingerprint Authentication System), consists of three main processing phases [3]: image acquisition, feature extraction, and matching. In the first phase, a sensor captures a fingerprint image, successively a features vector (containing information about minutiae) will be extract, and finally some indexes of similarity are calculated to verify the matching score between the input image and the stored template. In such system a scalar number called matching score is the considered measure of identity correctness. Generally: if the matching score is smaller than a fixed threshold the user will be refused otherwise the user will be accepted. 3.2

Fusion in the Proposed Multimodal Authentication System

A typical multimodal authentication system is composed of two or more parallel mono-modal systems and a fusion module. Fusion module takes two or more data and combines them in order to obtain the authentication result: impostor or genuine user. In this work a new module estimating input image quality have been added to a traditional AFAS. Image quality is estimated through the analysis of two specific image parameters: the number of erased segments (ES) and the number of the candidate minutiae (CM). The eliminated segments are portions of fingerprint too much noisy, so no useful information can be found inside them. The adopted matching algorithm is based on a structural matching executed on similar geometric forms between input fingerprint and the stored template. In particular triangle forms, produced by triplet of minutiae, have been used to perform the matching. We have called the produced score by a single AFAS “recognition degree”. The fusion strategies are divided into two main categories [6]: pre-mapping fusion (before matching) and post-mapping fusion (after matching). The first strategy deals with the sensor data fusion level and feature vector fusion level. These techniques are not used because they give many implementation problems [6]. The second strategy is realized through the decision level fusion, based on some algorithms which combine single decisions for each component system, or through the matching score level fusion, which combines the matching scores of each component system. Figure 1 shows the blocks scheme of a recognition process with many fusion levels.

4

V. Conti at al.

Fig. 1. Distinction of the different fusion levels: on the left the pre_mapping fusion, on the right the post_mapping fusion (used in this work). In this paper the decision level fusion and the matching level fusion were implemented and evaluated using two fingerprints based subsystems as sample. The whole recognition system also incorporates the knowledge about the image quality. As started before, image quality is based on both ES and CM features. Experimental trials aimed to find a global image Goodness Index (GI) in term of ES and CM values were performed. Figure 2 shows the ES and CM ranges with relative membership function values and the membership function range of the GI. These ranges are experimentally obtained. The Goodness Index GI=f(ES, CM) is obtained applying “moment defuzzify function” [16]. The moment defuzzify function returns a defuzzified floating point value, which represents the defuzzified fuzzy set using the centre of gravity (or first moment of inertia) function. The centroid method is a good method to use in many processes since it tends to smooth out the fuzzy region [16]. Figure 3 shows three different fingerprint images with a different GI value. The image on the left is a fingerprint of good quality, GI=82.01, on the center a fingerprint of medium quality, GI=55.01, and on the right a fingerprint of bad quality, GI=0.

Fig. 2. ES, CM and GI ranges. On the left ES values, on the center CM values and on the right GI values.


5

Fig. 3. Two typical fingerprints of different quality. On the left a fingerprint image of good quality (GI=82.01), on the center a fingerprint image of medium quality (GI=55.01), while on the right a fingerprint image of bad quality (GI=0). 3.3

Fuzzy Fusion

Decision level fusion The authentication system is composed of two modules AFAS to realize a multimodal authentication system combining the obtained scores from the single systems, Figure 4. With more details, the proposed authentication system is composed of “AFAS1” subsystem that performs index fingerprint processing and recognition, and of “AFAS2” subsystem that performs the middle finger’s fingerprint recognition and processing. The proposed and implemented fusion module uses the fuzzy logic principles and methods to combine the scores products by two subsystem AFAS [11]. Fuzzy logic processes imprecise information like human thinking and it allows to obtain intermediate values between true and false, accepted and refused, by partial membership set. A fuzzy inference system composed of two input and one output variables has been implemented. The output represents the decision taken by whole system.

Fig. 4. The general scheme of the proposed system architecture with decision fusion. The fuzzy logic conditions are formulated by a group of sixteen fuzzy rules. Some of the used rules used with this index quality are the following: 1. if recognition degree of AFAS1 is high and recognition degree of AFAS2 is high then final decision is genuine user.

6

V. Conti at al.

2. if recognition acknowledgment degree of AFAS1 is low and recognition degree of AFAS2 is low then final decision is impostor user. 3. if recognition degree of AFAS1 is medium low and recognition degree of AFAS2 is medium low then final decision is indecision. etc… Matching score level fusion The proposed system architecture is composed of two AFAS modules and the fusion is realized combining the matching score of both AFAS and the quality measure of fingerprint images. The fusion has been obtained by a fuzzy system with four input and one output variables. The output variable represents the decision of the whole system. The fuzzy system uses the knowledge base built with above fuzzy rules. Each rule has the following common guidelines: − if the input images have good quality and fingerprints are very similar to the stored fingerprints then certainly the user is who he/she claims to be. − if the input images have good quality and fingerprints are not very similar to the stored fingerprints then certainly the user is an impostor”. − if AFAS1 works with an input bad quality image and AFAS2 works with an input good quality image then the AFAS2 decision will be more discriminant than AFAS1. The figure 5 shows the scheme of the proposed architecture with matching score level fusion.

Fig. 5. General scheme of the proposed architecture with matching score level fusion.


4

7

Experimental Results

The purpose of an authentication system is to decide if a person is who he/she says to be. In other words, the system can make two types of errors: a False Rejection Rate (FRR, when a client is rejected) and a False Acceptance Rate (FAR, when an impostor is accepted) [13]. The performances of an authentication system are given in terms of error rates computed during the test phase. An optimal identity verification have FAR=0 and FRR=0. The overall performance of an identity verification system is well characterized by the Receiver Operating Characteristic (ROC) curve [5][13], which represents the FAR as a function of FRR or its complementary GAR=1-FRR (Genuine Acceptance Rate). The Equal Error Rate (EER: i.e. when FAR=FRR), is often used as the only performance measure of an identity verification method, although this measure gives just one point of the ROC. The algorithms are described in previous sections are used to merge the results coming from two modalities. The fuzzy rules have been deduced by analysis of the experimental trials. To evaluate algorithm performance the DB3 database of the FVC2002 [15] has been used. The database is composed of eight samples coming from one hundred users. The 800 images have 300*300 pixel dimension. The experimental results have been performed using the method of FVC competition to calculate FAR and FRR indexes [15]. The results show that the multimodal system performances are better than monomodal system. In addition, the tests show that the matching score level fusion gives better result than the decision level fusion ones. Figure 6 shows the performance of the proposed approach on DB3 using the ROC curves. In this figure the ROC curve of the mono-modal system without filter [4] is depicted. The Equal Error Rate is about 13,7% for decision level fusion and about 11,8% for matching score level fusion against 18,5% for mono-modal system. These experimental results show that using the matching score level fusion an improvement of 6.7% respect the mono-modal is obtained.

Fig. 6. The ROC curves about our approach on FVC DB.

8

5

V. Conti at al.

Conclusion

The multimodal methods for biometric authentication are used to overcome some limitations of a single modal system. In this work a multimodal biometric system which combines two different fingerprints to realize a personal authentication has been proposed. In particular, two type of fusion based on fuzzy logic techniques have been developed and tested. The experimental results show that the matching score level fusion gives good result. However, the system needs to be tested on a large database in a real operating environment.

References 1. V. Conti, G. Pilato, S. Vitabile, F. Sorbello, “A Robust System for Fingerprints Identification”, Knowledge-Based Intelligent Information Engineering System & Allied Technologies, Crema 16-18 September 2002, pp. 1162-1166. 2. V. Conti, G. Pilato, S. Vitabile, F. Sorbello, “Verification of Ink-on-paper Fingerprints by Using Image Processing Techniques and a New Matching Operator”, AI*IA Siena 10-13 September 2002, pp. 594 – 601. 3. V. Conti, G. Milici, S. Vitabile, F. Sorbello, “A New Directional Morphological Filter in a Automated Fingerprint Identification System”, accepted to CISI’06 Conferenza Italiana Sistemi Intelligenti, Ancona (Italia), September 17 - 29, 2006. 4. G. Milici, G. Raia, S. Vitabile, F. Sorbello, “Fingerprint Image Enhancement Using Morphological Filter”, IEEE EUROCON 2005 - The 8th International Conference on Computer as a tool. Belgrade, Serbia & Montenegro 21-24 November 2005. (pp. 967-970). 5. A. K. Jain, A. Ross, and S. Prabhakar, “An Introduction to Biometric Recognition ”, IEEE Transactions on Circuits and Systems for Video Technology, Vol.14, NO. 1, January 2004. 6. U. M. Bubeck, “Multibiometric Authentication – An Overview of Recent Developments” San Diego University, spring 2003, www.thuktun.org/cs574/papers/multibiometrics.pdf. 7. N. Poh, S. Bengio, and J. Korczak, "A multi-sample multi-source model for biometric authentication," in Proc. IEEE 12th Workshop on Neural Networks for Signal Processing, 2002, pp. 375--384. URL: http://citeseer.ist.psu.edu/thian02multisample.html. 8. A.K. Jain, L. Hong, Y. Kulkarni, “A Multimodal Biometric System Using Fingerprint, Face and Speech”, Conference on Audio-Video based Biometric Person Authentication 1999. 9. C. W. Lau, B. Ma, H. M. Meng, Y. S. Moon, Y. Yam, “Fuzzy Logic Decision Fusion in a Multimodal Biometric System”, Proceedings of the 8th International Conference on Spoken Language Processing (ICSLP), Korea, October 2004. 10. R. Zewail, A. Elsafi, M. Saeb, N. Hamdy, “Soft and Hard Biometrics Fusion for Improved Identity Verification”, the 47th IEEE International Midwest Symposium on Circuits and Systems 2004. 11. S. Prabhakar, A. K. Jain, “Decision-level Fusion in Biometric Verification”, Pattern Recognition, Vol. 35 (4), 2002, pp. 861-874. 12. A. K. Jain, L. Hong, and R. Bolle, “On-line fingerprint verification,”IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 4, pp. 302–313, 1997. 13. K. Dahel, Q. Xiao, “Accuracy Performance Analysis of Multimodal Biometrics”, Proceedings of the 2003 IEEE, Workshop on Information Assurance. 14. P. Verlinde, G. Chollet, and M. Acheroy, “Multi-Modal Identity Verification Using Expert Fusion”, Information Fusion, 1(1):17-33, July 2000. 15. http://bias.csr.unibo.it/fvc2002/ 16. L. A. Zadeh, “Fuzzy sets”, Information and Control 8, 338-353 (1965)