Efficient Data Hiding Techniques for Digital Rights ...... of-the-art multimedia production equipments (such as palm tops, digital camera ...... Surveys and Tutorials:.
Efficient Data Hiding Techniques for Digital Rights Management of Multimedia Archives
BY Hafiz Muhammad Aslam Malik B.S. (University of Engineering and Technology Lahore, Pakistan) 1999
Preliminary Proposal Submitted in partial fulfillment of the requirements for the degree of Ph.D. in the Graduate College of the University of Illinois at Chicago, 2004
Chicago, Illinois
TABLE OF CONTENTS CHAPTER 1.................................................................................................................................................................5 INTRODUCTION .........................................................................................................................................................5 MOTIVATION: ....................................................................................................ERROR! BOOKMARK NOT DEFINED. PROBLEM STATEMENT: ...............................................................................................................................................9 CHAPTER 2...............................................................................................................................................................12 RELATED WORK.......................................................................................ERROR! BOOKMARK NOT DEFINED. 2.1 DATA HIDING SYSTEMS: APPLICATIONS AND REQUIREMENTS.....................................................................12 2.1.1 REQUIREMENTS OF A DATA HIDING SYSTEM:...............................................................................................12 I. Robustness:...............................................................................................................................................13 II. Effectiveness:............................................................................................................................................13 III. Fidelity: ....................................................................................................................................................14 IV. Capacity: ..................................................................................................................................................14 V. Blind or Informed Detection: ...................................................................................................................14 VI. False Positive Rate:..................................................................................................................................14 VII. Multiple Watermarks Capability: ............................................................................................................15 VIII. Cost: ........................................................................................................................................................15 2.1.2 APPLICATIONS OF DATA HIDING FOR DIGITAL RIGHTS MANAGEMENT: ........................................................15 I. Ownership Protection: .............................................................................................................................15 II. Content Authentication:............................................................................................................................16 III. Fingerprinting:.........................................................................................................................................16 IV. Copy Protection: ......................................................................................................................................16 V. Broadcast Monitoring: .............................................................................................................................16 2.2. CLASSIFICATION OF DATA HIDING TECHNIQUES .......................................................................................17 2.2.1 CLASSIFICATION BASED ON HOST MEDIA TYPE............................................................................................18 I. Data Hiding in Images .............................................................................................................................18 II. Data Hiding in Video................................................................................................................................18 III. Data Hiding in Audio ...............................................................................................................................18 IV. Data Hiding in Text ..................................................................................................................................18 2.2.2 CLASSIFICATION BASED ON DATA HIDING APPLICATIONS ............................................................................18 I. Robust Data Hiding..................................................................................................................................18 II. Fragile Data Hiding .................................................................................................................................18 III. Semi-Fragile Data Hiding........................................................................................................................18 2.2.3 CLASSIFICATION BASED ON PERCEPTIBILITY................................................................................................18 I. Imperceptible Data Embedding................................................................................................................19 II. Visible Data Embedding...........................................................................................................................19 2.2.4 CLASSIFICATION BASED ON DATA EMBEDDING DOMAIN ..............................................................................19 I. Data Hiding in Spatial/Time Domain (Direct Domain) ...........................................................................19 II. Data Hiding in Transformed Domain.......................................................................................................19 2.2.5 CLASSIFICATION BASED ON DATA EMBEDDING METHOD .............................................................................20 I. Additive Spread Spectrum or Host-Interference-Non-rejecting Methods.................................................20 II. Host Interference Rejecting Methods .......................................................................................................20 2.2.6 CLASSIFICATION BASED ON DATA EXTRACTION METHOD .............................................................................20 I. Private or Informed Data Hiding .............................................................................................................21 II. Semi-Private Data Hiding ........................................................................................................................21 III. Public or Blind Data Hiding ....................................................................................................................21 3.3 DIGITAL RIGHTS MANAGEMENT: A BRIEF OVERVIEW ................................................................................23 CHAPTER SUMMERY ..................................................................................................................................................25 CHAPTER 3...............................................................................................................................................................26
2
DATA HIDING MODELS...........................................................................ERROR! BOOKMARK NOT DEFINED. 3.1 3.2 3.2.1 3.2.2 3.2.3 3.3 3.3.1 3.3.2 3.3.3 3.4
NOTATION ...........................................................................................ERROR! BOOKMARK NOT DEFINED. TRANSMISSION CHANNELS ........................................................................................................................21 BOUNDED DISTORTION CHANNELS..............................................................................................................21 BOUNDED HOST-DISTORTION CHANNELS ....................................................................................................22 ADDITIVE NOISE CHANNELS ........................................................................................................................22 DATA HIDING IN COMMUNICATION FRAMEWORK ...............................ERROR! BOOKMARK NOT DEFINED. CLASSICAL MODEL OF COMMUNICATIONS SYSTEM .................................ERROR! BOOKMARK NOT DEFINED. SECURE TRANSMISSION .........................................................................ERROR! BOOKMARK NOT DEFINED. DATA HIDING MODEL BASED ON COMMUNICATION...............................ERROR! BOOKMARK NOT DEFINED. DATA HIDING AS COMMUNICATION WITH SIDE INFORMATION AT THE TRANSMITTER ERROR! BOOKMARK NOT DEFINED. 3.5 GEOMETRIC MODEL OF DATA HIDING ................................................ERROR! BOOKMARK NOT DEFINED. CHAPTER SUMMERY ............................................................................................ERROR! BOOKMARK NOT DEFINED. CHAPTER 4...............................................................................................................................................................44 BLIND DATA EMBEDDING ....................................................................................................................................44 4.1 DATA HIDING BASED ON ADDITIVE EMBEDDING ................................ERROR! BOOKMARK NOT DEFINED. 4.2 WORK IN PROGRESS: ROBUST AND HIGH RATE DATA EMBEDDING .............................................................44 4.2.1 DATA HIDING USING FREQUENCY SELECTIVE BASED SPREAD SPECTRUM ....................................................44 4.2.1.1 WATERMARKING USING PERCEPTUAL AUDITORY MODEL .................................................45 4.2.1.2 SALIENT POINT EXTRACTION ....................................................................................................46 4.2.1.3 WATERMARK EMBEDDING.........................................................................................................48 4.2.1.3.1 Watermark Generation ...................................................................................................................49 4.2.1.3.2 Watermark Embedding ...................................................................................................................49 4.2.1.4 WATERMARK DETECTION ..........................................................................................................51 4.2.1.5 EXPERIMENTAL RESULTS...........................................................................................................52 4.3 FUTURE DIRECTIONS .................................................................................................................................55 4.3.1 PROPOSED DADA HIDING SCHEME FOR IMAGES ..........................................................................................55 4.3.2 PROPOSED DADA HIDING SCHEME FOR VIDEO ...........................................................................................57 CHAPTER 5...............................................................................................................................................................59 INFORMED DATA EMBEDDING ...........................................................................................................................59 5.1 INFORMED EMBEDDING .......................................................................ERROR! BOOKMARK NOT DEFINED. 5.1.1 COSTA’S WORK ....................................................................................ERROR! BOOKMARK NOT DEFINED. 5.2 QUANTIZATION INDEX MODULATION (QIM) ......................................ERROR! BOOKMARK NOT DEFINED. 5.2.1 BINARY DITHER MODULATION ..............................................................ERROR! BOOKMARK NOT DEFINED. 5.3 WORK IN PROGRESS: HIGH RATE DATA EMBEDDING USING INFORMED ENCODING ....................................59 5.3.1 DATA HIDING USING FREQUENCY SELECTIVE DITHERING (OUR CONTRIBUTION).........................................59 5.3.1.1 FIR APPROXIMATION OF APF....................................................................................................60 5.3.1.2 DATA EMBEDDING ......................................................................................................................63 5.3.1.3 DATA DETECTION USING SIGNAL MODELING .......................................................................64 5.3.1.3.1 Spectrum Estimation .......................................................................................................................65 5.3.1.3.2 Allpass Filter Parameter Estimation ..............................................................................................66 5.3.1.3.3 Simulation Results...........................................................................................................................67 5.3.1.4 DATA DETECTION USING MATCH FILTER...............................................................................68 5.3.1.4.1 Simulation Results...........................................................................................................................70 5.4 FUTURE DIRECTION ...................................................................................................................................71 5.4.1 EXTENSION AUDIO FINGERPRINTING AND AUTHENTICATION ........................................................................72 CHAPTER 6...............................................................................................................................................................73 CONCLUSION & FUTURE DIRECTIONS ..............................................................................................................73 REFERENCES: .........................................................................................................................................................75
3
TABLE OF FIGURES FIGURE 2.1: GENERAL CLASSIFICATION OF DATA HIDING ...........................................................................................17 FIGURE 2.2: ANATOMY OF A DRM TRANSATION ........................................................................................................21 FIGURE 3.1: STANDARD COMMUNICATION MODEL ................................................ERROR! BOOKMARK NOT DEFINED. FIGURE 3.2: STANDARD SECURE COMMUNICATION MODEL ..................................ERROR! BOOKMARK NOT DEFINED. FIGURE 3.3: GENERAL MODEL FOR DATA HIDING .................................................ERROR! BOOKMARK NOT DEFINED. FIGURE 3.4: DATA HIDING SYSTEM WITH INFORMED DETECTOR ANALOGOUS TO STANDARD SECURE COMMUNICATION MODEL............................................................................ERROR! BOOKMARK NOT DEFINED. FIGURE 3.5: DATA HIDING SYSTEM WITH BLIND DETECTOR ANALOGOUS TO STANDARD SECURE COMMUNICATION MODEL ............................................................................ERROR! BOOKMARK NOT DEFINED. FIGURE 3.6: DATA HIDING AS COMMUNICATION WITH SIDE INFORMATION AT THE ENCODER...... ERROR! BOOKMARK NOT DEFINED. FIGURE 4.1: 5 –LEVEL MODIFIED DISCRETE WAVELET ANALYSIS FILTER BANK ........................................................48 FIGURE 4.2: BLOCK DIAGRAM OF WATERMARK EMBEDDING PROCESS .......................................................................50 FIGURE 4.3: NORMALIZED CORRELATION FOR WATERMARKED SUBBAND (LEFT) AND UNWATERMARKED SUBBAND (RIGHT). ..............................................................................................................................................52 FIGURE 4.4: BLOCK DIAGRAM FOR WATERMARK DETECTION .....................................................................................52 FIGURE 4.5: DPM FOR DIFFERENT VALUES OF NOISE POWER (PN).................................................................................54 FIGURE 5.1: INFORMED DATA EMBEDDING FOLLOWED BY AWGN ATTACK .........ERROR! BOOKMARK NOT DEFINED. FIGURE 5.2: BINARY DITHERED MODULATION SCHEME BASED ON DITHERED UNIFORM SCALAR QUANTIZATION ............................................................................................ERROR! BOOKMARK NOT DEFINED. FIGURE 5.3: MAGNITUDE RESPONSE OF APF H(EJW) APPROXIMATION FOR DIFFERENT VALUES OF LENGTH (L). ........62 FIGURE5.4: POLE-ZERO LAYOUT OF HAPI(Z) FOR BINARY ENCODING ..........................................................................62 FIGURE5.5: POLE-ZERO LAYOUT OF HAPI(Z) FOR 4-ARY ENCODING .............................................................................63 FIGURE 5.6: BLOCK DIAGRAM OF THE DATA EMBEDDING SCHEME .............................................................................64 FIGURE 5.7: BLOCK DIAGRAM OF THE DATA DETECTION PROCESS .............................................................................67 FIGURE 5.8: PROBABILITY OF ERROR (PE) VS. SNR PLOT FOR BOTH ENCODING SCHEMES ..........................................67 FIGURE 5.9: MAGNITUDE SPECTRUM OF CZT OF THE SUBBAND SEQUENCE X4,6(N) BEFORE AND AFTER PASSING THROUGH H0(Z I) I.E. Y4,6(N),AT R = 0.9 (RIGHT) AND AT R = 1/0.9 (RIGHT). .........................................69 FIGURE 5.10: BLOCK DIAGRAM OF THE DATA DETECTION USING MATCH FILTER ......................................................70 FIGURE 5.11: PROBABILITY OF ERROR FOR DIFFERENT SNR VALUES...........................................................................71
4
CHAPTER 1
Introduction The revolution in the area of digital information has visibly impacted our society and everyday life [171, 172]. Some of the blessings of this digital revolution include: the evaluation of Internet as a global village, availability of low–cost large capacity storage devices, deployment of long– distance seamless networks at Gbps (gigabits per second) data rates, and popular use of the stateof-the-art multimedia production equipments (such as palm tops, digital camera, camcorder, high-tech scanner and printer, digital audio recorder, etc.). Furthermore, the developments in the areas of digital media production, manipulation, and distribution have added new dimensions to the technical challenges related to digital data security and integrity. Along with its countless advantages the cutting edge technologies of this digital information revolution have generated some serious concerns about digital content protection, ownership protection, unauthorized copy prevention, etc. Today’s entertainment industry (music and film industry) alone claims a multimillion dollar annual revenue loss due to piracy [171], which is more likely to increase in the coming years due to fast growing trend of exchanging digital media (music, images, movies, software, e-books, etc.) over peer-to-peer networks. There is an urgent need to develop robust technologies to support the development of digital rights management (DRM) systems, capable of providing diverse services such as, secure media streaming between user and content server, ownership protection, unauthorized copy prevention, unauthorized content usage, content authentication, and content usage tracing. Generally digital rights management (DRM) systems consist of a set of rights models (business models) and technologies to support the above-mentioned services. However, the research proposed in this dissertation deals only with some of the technological issues of DRM systems. These technological issues define the reliability of a DRM system.
5
Most of the existing DRM systems use traditional content protection schemes such as encryption for digital content protection, secure content delivery, and its usage tracking [170]. However, encryption and scrambling alone cannot provide adequate protection against ownership rights, unauthorized content usage, unauthorized copy prevention etc. Encrypted or scrambled data remain protected as long as decryption or unscrambling key is unknown, but once data is decrypted or unscrambled there is no way to stop its reproduction or sharing [168]. Thus there is a strong need to complement cryptography. Data hiding and watermarking (a special case of data hiding) are the potential technologies promising to meet the shortcomings of traditional contentprotection technologies. In general, information hiding or data hiding implies imperceptibility embedding information (message or metadata) into the host signal (images, video, audio, text etc.) for a variety of applications such as secret communication or steganography, content protection, ownership protection, illegal copy prevention, etc. Salient characteristics of any data hiding scheme include: embedding capacity or payload, minimal embedding distortion, robustness to attacks, low false positive rate, low error probability of received data, etc. Among these, embedding capacity, embedding distortion or fidelity, and robustness are three inter-dependent features, and are also used to evaluate the performance of data embedding schemes. Embedding capacity refers to the amount of data that can be embedded in a give multimedia clip. Embedding distortion or fidelity measures the perceptibility of the embedded information. Robustness refers to the capability of data hiding scheme to withstand intentional and unintentional attacks. Here, intentional attacks include filtering, chopping, scaling, Gaussian or uniform noise addition, resampling, etc., whereas, lossy compression digital to analog conversion and requantization are generally treated as unintentional attacks. Digital watermarking, a special case of data hiding, is a process of embedding information into the host data (cover data) for content protection, integrity and security. Robustness of the 6
embedded information against data hiding attacks is the most desirable feature of watermarking schemes. A watermark is an imperceptible and inseparable signal about the data in which it is embedded, and undergoes same transformation as the host data. These attributes distinguish watermarking from the traditional digital content protection techniques [7] such as cryptography and scrambling and this make watermarking an attractive tool for digital media protection, traitor tracing, content usage monitoring, broadcast monitoring, and communication with side information to improve the quality of service (QoS) of the multimedia transmission over lossy channel. The growing availability of digital information in different formats and its increasing illegal sharing and distribution has led to the proliferation of DRM technologies including, data hiding schemes designed for applications such as copyrights protection, media authentication, broad cost monitoring [5 – 33], steganographic techniques for covert communications [8, 127 – 137], fingerprinting methods for traitor-tracing applications [90 – 99]. This has also led to renewed interests of information theoreticians in the data hiding problem, e.g. Moulin et al [39 – 41, 47 – 54], Cox et al [7, 59, 69, 71, 139, 130, 144, 145, 150], Chen et al [80 – 86], Girod et al [57, 58, 70, 104 – 112], P-Gonzalez et al [74 – 76, 115, 135, 146 – 149], Cohen et al [43 – 46, 55, 56], and others [6, 25 – 33]. Most of the theoretical advances in the area of data hiding are attributed to the following classical works: “Writing on Dirty Papers” by M. Costa [37] “Coding of Channels with Random Parameters” by Gel’fand and Pinsker [35], “Channels with Side Information at the Transmitter” by Shannon [34]. Due to these inspirational papers many researchers have modeled the data hiding problem using signal processing, communication theory, coding theory, and information theory framework. From an application perspective, most of the existing data hiding research [9 – 24] is mainly focused on digital images data. Relatively very little attention has been given to data hiding in 7
digital video and audio data. Audio data models (perceptual as well as real data models) are quite different from images and video data models. Therefore, a data hiding scheme yielding high performance of a given data hiding scheme for images or video may not yield the same performance for audio, and vice versa. In the following we outline various shortcomings of the existing data hiding schemes. More detailed analysis of the related work will be provided in Chapter 3. First of all, the host data is generally modeled as an independent and identically distributed (i.i.d.) Gaussian random sequence, and the attack channel is modeled as an independent discrete memoryless (DM) Gaussian channel [39 – 56]. These models do not agree with the host data (audio, video, and images) models, because, in general multimedia data does not exhibits i.i.d. Gaussian distribution [9]. Similarly, in practice active adversary attacks are host data dependent [104 – 126], especially when an active adversary has knowledge about the host data. Therefore, there is a need for more realistic and appropriate modeling of host data and attack channel for performance analysis of a given data-hiding scheme. Secondly, almost all existing data hiding schemes measure the perceptual quality (fidelity) of a given data hiding scheme using the mean squared error metric [39 – 41, 80 – 86, 115, 135, 146 – 149], which often does not agree with the human perceptual model [2, 7, 60]. An appropriate perceptual distortion measure is also needed for the performance evaluation based on the perceptual distortion due to information embedding and robustness. Thirdly, low data rate is a common limitation of the existing data hiding schemes [5 – 24]. Since data hiding applications such as broadcast monitoring require relatively high data rate, therefore, it is desirable to develop high capacity data hiding schemes for such applications. While several researchers [71 – 86] have proposed high capacity data hiding schemes, but their work is based on improving the coder performance by using efficient coding schemes and/or using host signal interference cancellation [77 – 89]. A little attention has been focused on exploiting the host data 8
characteristics combined with a efficient coder and data hiding strategy to achieve high data embedding rate. Moreover, most of the existing DRM systems use encryption for content protection and content tracking which alone cannot provide sufficient firewall against active adversary attacks. Finally, most of the research in the data hiding community is mainly focused on traditional copy control issues such as, copyright protection, content authentication, temper detection, unauthorized copy prevention, content usage monitoring, broadcast monitoring, etc. Very little attention has been given to broaden the data hiding application domain beyond the copy control issues. For example, data hiding can be used in the area of multimedia transmission over lossy and bursty channel to improve the QoS.
Problem Statement: Most of the existing data hiding schemes [4 – 33] are based on i.i.d. Gaussian modeling of the host data and independent Gaussian discrete memoryless channel (DMC) modeling of adversary attacks or attacks channels [39 – 56]. These assumptions are not true in general. Similarly, performance measures based on embedding fidelity of existing data embedding schemes generally use mean squared error distortion criterion which does not agree with the human audiovisual perceptual model. Therefore it is desirable to develop more realistic host data, attack channel, and embedding distortion models for performance analysis of existing data embedding schemes. The main goal of this research is to advance the theory underlying data hiding, develop new techniques for data hiding, and extend data hiding applications to digital rights management system. In this dissertation we propose to analyze the limitations of existing data hiding schemes and their applications to different types of host data. Based on the analysis we will propose efficient data hiding schemes for a reliable DRM system. We intend to analyze the performance of the proposed schemes based on the triad of data hiding performance criteria, that is, capacity, 9
perceptibility, and robustness using more realistic data and channel models. We also plan to develop a more realistic measure of distortion due to embedding in order to evaluate the perceptual performance of the proposed data-hiding scheme. In particular we propose to investigate the following research tasks:
Develop high capacity data hiding schemes based on host signal features along with efficient codec and data hiding strategies.
Devise realistic host data models (stochastic models) for each type of host media i.e. audio, images, and video separately for information-theoretic analysis of the proposed data hiding schemes.
Design efficient source coding schemes based on the developed host data models.
Develop appropriate channel models for intentional and unintentional attacks and analyze their performance against existing attack channel models.
Devise a realistic distortion metric to evaluate the performance based on perceptibility and robustness.
Develop an appropriate protocol for online multimedia authentication, copy control, and copyright protection applications.
Devise suitable data-hiding scheme for multimedia indexing and retrieval application.
Develop a realistic data hiding strategy to improve the QoS of multimedia transmission over lossy and busty channels.
The remainder of the dissertation proposal is organized as follows: Chapter 2 discusses the requirements and application domain of data hiding schemes along with a general classification of existing data hiding schemes. A brief overview of a DRM system is also provided in Chapter 2. Related work and data hiding modeling is given in Chapter 3. Our contribution to blind data embedding or additive spread spectrum class of data hiding is discussed in Chapter 4. Chapter 5 10
gives the details of our proposed work in informed embedding class of data hiding. Our proposed data hiding schemes for both classes of data hiding use audio data as a host media, their extensions for images and video host data are also proposed. Future directions of our proposed research are outlined in Chapter 6.
11
CHAPTER 2
Preliminaries This chapter presents an overview of the generic characteristics and requirements of the data hiding problem, briefly describes the related application domains, and provides a general classification of the existing data hiding schemes. The challenges, shortcomings, and the promises of the data hiding schemes are outlined in Section 2.1. Section 2.2 gives general classification of existing data hiding schemes. Transmission channel model is an important ingredient for theoretical analysis of data hiding problem. Common transmission channel models, such as bounded distortion channels, bounded host distortion channels, and additive noise channels that have been used for modeling attacks against data hiding schemes and watermarking are discussed in Section 2.3. A brief overview of DRM systems is provided in Section 2.4.
2.1
Data Hiding Systems: Requirements and Applications
The digital multimedia (throughout this document digital multimedia or the host media refers to digital audio, digital video and digital images, unless otherwise specified) has many advantages over analog multimedia. For example, there is insignificant aging effect on the contemporary digital media storage devices such as CDs, memory sticks, etc., reproduction of digital media is very simple; a copy of a digital media clip is exactly similar to its original version. Also due to recent advances in the techniques for digital data production, distribution, and manipulation, research in the area of data hiding and watermarking has exploded with the goal to complement deficiencies of the conventional content protection methods such as cryptography and scrambling [7, 8].
2.1.1 Requirements of a Data Hiding System: 12
A data hiding scheme is characterized by a number of defining properties [5 – 8]. In general a data hiding scheme is suppose to withstand against common data manipulations, such as lossy compression, digital-to-analog conversion, rescaling, requantization, resampling, filtering, data format conversion, encryption, decryption, and scrambling. It is also suppose to withstand against active adversary attacks, such as noise, as long as attack channel distortion is below a certain masking threshold. However, the relative importance of each property depends on the requirements of the application and the role of data embedding in the application. For example, if we are evaluating the performance of an audio watermarking system for copy control application, we may need to check the robustness of short time energy ratio that adversary might use for attack. However, such robustness might be irrelevant for broadcast monitoring applications. Therefore, the performance of any data hiding scheme should be evaluated based on the underlying application. Following are the desirable properties of a generic data hiding scheme:
I.
Robustness:
Robustness measures the ability of embedded data or watermark to withstand against intentional and unintentional attacks. Unintentional attacks generally include common data processing operations i.e. compression, digital-to-analog conversion, resampling, requantization etc, where as, intentional attacks cover a broad range of degradations [104 – 126], for example, white and color noise addition, scaling, rotation (for image and video watermarking schemes), chopping, low-pass filtering, etc. Details of these intentional attacks in the area of data hiding and their countermeasures can be found in [8, 131, 132].
II.
Effectiveness:
The probability that the output of the embedder will be watermarked for a randomly selected input data is generally referred as effectiveness of a data hiding scheme.
13
III.
Fidelity:
This is an important property of all perceptual based data hiding schemes [5 – 24]. Fidelity measures the perceptual similarity between the host media and its data embedded version. To meet this constraint, the perceptual distortion introduced due to embedding is kept below the masking threshold of human auditory system (HAS) for audio data hiding schemes and human visual system (HVS) for video and image data hiding schemes.
IV.
Capacity:
This property refers to the amount of information that a data hiding scheme can successfully embed without introducing perceptual distortion. The need for this property is application dependent, for example, a data hiding scheme designed for copyright protection or copy control application does not require high data embedding capacity because only a few bits of information are sufficient for this application. Whereas, a data embedding scheme for broadcast monitoring applications requires to embed relatively large amount of data [6, 7].
V.
Blind or Informed Detection:
This property relates to the availability of host data at the detector for watermark detection process. If the host data is available at the detector for watermark detection process; then, this class of data hiding schemes are categorized as informed detector or private data hiding schemes. These schemes are required for fingerprinting, and data authentication [5 – 7]. If the host data is not available at the detector for watermark detection process then this class of data hiding schemes are categorized as blind detector or public data hiding schemes. Blind detector based data hiding schemes are commonly used for copy control applications.
VI.
False Positive Rate:
This property corresponds to the frequency of detecting mark in an unmarked portion of the host data. It is an important property for content protection applications such as, ownership right, copy control, etc.
14
VII.
Multiple Watermarks Capability:
This feature of a data hiding scheme to embed more than one mark in the same host data is desirable in some application such as fingerprinting.. For example, consider a situation where the owner and the chain of distributors of a multimedia product want to embed their marks (serial numbers or tags) to keep the trace of content usage and tracing a traitor. For such applications multiple watermarks embedding feature is desirable.
VIII.
Cost:
The computational cost of embedding and detection algorithm is another evaluation criterion of data hiding schemes that is critical for real time applications, such as broadcast monitoring, online content authentication, etc. On the otherhand, for ownership proof applications this property is not that critical.
2.1.2 Applications of Data Hiding: Applications domain of data hiding techniques is rapidly growing. Recently, several research efforts [5, 9, 10, 150 – 158] are aiming beyond classic applications of data hiding including ownership protection, content usage tracking, content authentication, copy control, fingerprinting, broadcast monitoring, indexing, medical safety [5 – 24] etc. A brief overview of a few of these applications and their design requirements is given in the following:
I.
Ownership Protection:
The watermark carrying the ownership information is embedded into the host data. The watermarking scheme used for ownership protection is expected to be resilient to common data processing operations (unintentional attacks) and intentional attacks. In the case of dispute over ownership of the host data, embedded watermark can be used as a proof to identify the true owner of the host data. Watermarking schemes intended for ownership protection must have low probability of error and false alarm. In general, the capacity (payload) requirement of the watermarking scheme designed for ownership protection applications does not need to be high.
15
II.
Content Authentication:
Robustness and undetectability are not the main concerns for content authentication application of data hiding. Therefore, fragile watermarking is generally used for such applications. A watermark is embedded in the host data, which is later used to determine the tempering of the host media. Recent content authentications schemes are also capable of identifying the locations of tempering in the host media [9 – 24, 100 – 103]. These applications generally require informed detector i.e. original host data is available to the detector for content authentication. Data hiding schemes for content authentication must have high embedding capacity to meet the requirements of the content authentication applications [6].
III.
Fingerprinting:
The owner or distributor of multimedia contents uses fingerprinting or labeling to trace the illegal copies or traitor. For such applications, content owner or distributor embed a unique fingerprint, label, or serial number in each copy of the distributed data before distributing to each customer. A fingerprinting scheme is required to survive against intentional and unintentional attacks, more specifically collusion attacks [90 – 99]. Fingerprinting does not require high embedding capacity but does require robustness in general.
IV.
Copy Protection:
Embedded information in the host multimedia data can be used to control the copying device for unauthorized copy prevention [7]. For this purpose, a watermark detector is generally integrated in the recording or playback system, such as, DVD copy control scheme proposed in [150], or proposed SMDI player [159]. Data hiding schemes for such applications should be robust against all intentional or unintentional attacks that temper with the watermark from the watermarked data. Moreover, data hiding techniques designed for copy control intend to use a blind detector and generally requires low data embedding capacity.
V.
Broadcast Monitoring:
An automated (active) broadcast monitoring system can be used to detect the embedded watermark in the broadcasted commercial advertisement [5, 7, 158]. In addition, an active 16
broadcast surveillance system can also be used for other TV products (news, talk shows, etc.) protected by broadcast monitoring watermarking systems. For such applications watermarking scheme should be robust against watermarking attacks and requires a blind detector for watermark detection process. Furthermore, such applications require low watermark embedding capacity.
2.2.
Classification of Data Hiding Techniques
This section provides a general classification of existing data hiding techniques based on the following six criteria:
host media type (images, video, audio, and text),
areas of applications (robust, fragile, and semi-fragile),
perceptibility (visible and invisible),
embedding domain (spatial and transform),
data embedding schemes (know-host-state and know-host-statistics), and
data extraction techniques (private, semi-private, and public).
This classification hierarchy of data hiding techniques is illustrated in Figure 2.1. DATA HIDING
BASED ON HOST MEDIA TYPE
BASED ON APPLICATIONS
BASED ON PERCEPTIBILITY
IMAGE DATA HIDING
ROBUST DATA HIDING
IMPERCEPTIBLE DATA HIDING
DIRECT DOMAIN EMBEDDING
VIDEO DATA HIDING
SEMI-FRAGILE DATA HIDING
VISIBLE DATA HIDING
TRANSFORMED DOMAIN EMBEDDING
AUDIO DATA HIDING
FRAGILE DATA HIDING
TEXT DATA HIDING
BASED ON EMBEDDING DOMAIN
BASED ON EMBEDDING SCHEME HOST INTERFERENCE CANCELLATION TECHNIQUES (INFORMED DATA EMBEDDING) ADDITIVE SPREAD SPECTRUM TECHNIQUES (BLIND DATA EMBEDDING)
BASED ON EXTRACTION SCHEME
PRIVATE DATA HIDING SEMI-PRIVATE DATA HIDING PUBLIC DATA HIDING
Figure 2.1: General Classification of Data Hiding
17
Each category of the data hiding schemes is discussed briefly as follow,
2.2.1 Classification Based on Host Media Type Most of the data hiding research is focused on digital images compared with the other host media types i.e. video, audio, and text. This is due to the fact that the performance evaluation of a data hiding scheme for digital images is relatively easier than digital audio and video; because the performance evaluation of a data embedding scheme for audio or video generally requires subjective testing. Data hiding techniques based on host media type can be divided into four subgroups [127 – 150]: 1. 2. 3. 4.
Data Hiding in Images Data Hiding in Video Data Hiding in Audio Data Hiding in Text
2.2.2 Classification Based on Data Hiding Applications Performance based on robustness, capacity and fidelity of a data hiding scheme depends on the application of interest. For example, copyrights protection applications require a robust watermarking [57 – 89] where as content verification applications need a fragile watermarking [7, 9 – 24]. Similarly, fingerprinting needs a semi-fragile watermarking [90 – 103]. Therefore, existing data hiding schemes can be classified into three sub-groups based on the application of interest: 1. Robust Data Hiding 2. Fragile Data Hiding 3. Semi-Fragile Data Hiding
2.2.3 Classification Based on Perceptibility Existing data hiding schemes can be divided into two main categories based on the perceptibility (fidelity) of embedded data [5 – 24], that is, 18
1. Imperceptible Data Embedding 2. Visible Data Embedding Imperceptible data embedding implies that embedded data is invisible (in case of image, video, and text host media) and inaudible (for audio host media). Imperceptible data embedding schemes are more common than the visible data embedding schemes [60 – 68]. Imperceptible data embedding schemes exploit the HVS and HAS characteristics to ensure imperceptibility of the embedded data. Visible data embedding schemes are generally used to imprint visible logo in digital images or video.
2.2.4 Classification Based on Data Embedding Domain Existing data hiding schemes can be classified into two major categories based on embedding domain of the host media, that is, 1. Data Hiding in Spatial/Time Domain (Direct Domain) 2. Data Hiding in Transformed Domain Least significant bit (LSB) encoding, patchwork, echo hiding, etc. are few common data hiding schemes of direct domain data embedding [127, 128, 131, 136, 160] schemes. Direct domain data hiding schemes very popular among the data hiding community. Discrete cosine transform (DCT), discrete wavelet transform (DWT), and discrete fourier transform (DFT) are the most commonly used transforms for data embedding process. Most DCT-based image data embedding methods commonly use 8x8 size block of image for host data transformation then watermark is embedded by modifying DCT-coefficients according to HVM [9 – 24]. In DWT-based data embedding algorithms the host data is first decomposed into subbands using DWT, then for data embedding discrete wavelet coefficients in the selected subbands are modified based on human perceptual model. Robust data hiding schemes for images and video resilient to rotational, scaling, and translational (RST) distortion generally use DFT-based data hiding schemes [6, 7, 127 – 141, 143, 144]. DFT-based algorithms are also common for audio data hiding schemes. 19
2.2.5 Classification Based on Data Embedding Method Existing data hiding schemes based on the data embedding methods can be classified into two major categories [75 – 86, 115, 139], that is, 1. Additive Spread Spectrum or Host-Interference-Non-Rejecting Methods 2. Host Interference Rejecting Methods or Informed Embedding In case of additive spread spectrum based data hiding, a pseudorandom sequence w(mi) generated using secret key or message mi is added to the host signal i.e. x ( m i ) = C o +α × w ( m i )
2 .1
here 0 < α ≤ 1 where α is called as scaling factor and value α is the tradeoff between robustness and fidelity of the embedded data. From Eq. 2.1 this is clear that for this class of data hiding the host signal Co acts as an additive interference if a blind detector is used for watermark detection, which ultimately limits the performance of the detector; even in the absence of attack channel zero-error probability is hard to achieve. But these methods out perform the host interference rejection methods under sever attack situations. Most of the existing data hiding methods [9 – 24] fall into the additive spread spectrum class. The inherited limitations of the host interference non-rejecting methods can be improved by exploiting the host signal knowledge at the encoder; these methods are generally known as host interference rejecting methods. Quantization index modulation (QIM) [77 – 89] based data hiding methods is a sub-class of host interference rejecting methods. This class of data hiding methods provides an easy control over the trade off between data rate, embedding distortion, and robustness. These methods generally have higher data rate than the spread spectrum based data hiding class at the cost complexity of the data hiding system.
2.2.6 Classification Based on Data Extraction Method 20
Data hiding systems based on the information available at detector can be classified in following categories, 1. Private or Informed Data Hiding 2. Semi-Private Data Hiding 3. Public or Blind Data Hiding Private data hiding systems requires original copy of the host media along with secret embedding key for data extraction. These systems are generally used for data hiding applications like content authentication, ownership verification, etc. Semi-private systems generally requires secret embedding key only for information extraction, whereas, public data hiding systems need only a marked copy of the host media at the detector for data extraction [5 – 33].
2.3
Transmission Channels
The transmission channel model plays an important role in analyzing the performance of a given communications system. In general, a fixed transmission channel is assumed for design and analysis of a communication system i.e. we cannot modify or design the noise function that occurs during transmission. A channel is generally characterized by means of a conditional probability distribution, Pr/x (r/x) which gives the probability of obtaining r at the output of the channel when x is the input of the channel. Transmission channels are modeled based on the noise function they apply to the signal and how the noise is applied. In data hiding scenario adversary attacks (or attack channel) are generally treated as a transmission channel for the performance analysis of a data hiding scheme. Commonly known attacks in the data hiding community can be modeled as follow [30, 115]:
2.3.1 Bounded Distortion Channels In this case we consider the largest distortion energy per dimension σ n2 to ensure mˆ = m (zeroerror) for any distortion (noise) vector n, that satisfies,
21
n
2
≤ N σ n2
2.2
This channel model describes the minimum signal to noise ration (SNR) constraint between the attack channel input and output. Bounded distortion channel model is more appropriate for unintentional attacks such as compression attacks or active adversary attack to remove watermark for the watermarked media.
2.3.2 Bounded Host-Distortion Channels Some active adversary may use distortion measure between the host signals instead of distortion introduced by channel. Since this is a direct measure of degradation of the host signal. This model is appropriate when an attacker has partial knowledge about the host signal, this might be in probabilistic sense i.e. probability distribution of the host signal is know or any other sence. Active adversary can calculate the distortion between a watermarked copy of media and the host media, this distortion is bounded to the expected distortion given as D r = E [ D ( r , x )]
2.3
where expectation is taken over the conditional probability density of r given the channel input x.
2.3.3 Additive Noise Channels In this case noise vector n is modeled as random and statistically independent of the host data Do [39 – 46]. An additive white Gaussian noise (AWGN) is an example of such channel. The robustness measure in this case is the maximum noise variance σ n2 to ensure sufficiently low probability of error in the received data. Many researchers in the area of data hiding use AWGN channel assumption to model attack channel for performance analysis of a given data hiding scheme [5 – 7, 25 – 33].
22
The first two channel models are distortion constraints which are more appropriate to model intentional attacks [5, 25 – 33, 115, 139] whereas AWGN channel is appropriate for unintentional attacks.
3.4
Digital Rights Management: A Brief Overview
Digital rights management (DRM), i.e., the technologies, tools, and processes that protect intellectual property during the life cycle of digital content, is a vital ingredient of the emerging electronic multimedia (emedia) market. DRM creates an essential foundation of trust between authors and consumers that is a prerequisite for the robust market development. At its simplest level, digital rights management (DRM) technology is all about controlling access to information. Customers want convenient access to their purchased products, while companies seek to protect their intellectual property from unauthorized use or duplication. DRM sits squarely between these two parties, trying to present an amicable compromise between the customers and the vendor. The hardware keys, software licenses, and serial numbers all fall under the DRM umbrella [169]. Although there are several approaches to providing digital rights management, but "Anatomy of a DRM Transaction" is the most common one which is outlined in Figure 2.2. 3 PLAY MANAGEMENT SYSTEM
4 CLIENT W EB BROWSER
1
2
CONTENT AUTHOR/ CREATOR
MEDIA CONVERTER
W EB STOREFRONT AND MEDIA HOST
6 CONSUMER
5 CLIENT VIEWER
LICENSE MANAGER
Content Manager
Figure 2.2 Anatomy of a DRM Transaction
23
In Figure 2.2, at its most basic level, a DRM transaction starts with the content creator (1), who generates a piece of media (2), be it audio, video, text, or some other format. Once in digital form, the media file is encrypted or watermarked to protect it from unauthorized use and stored on a content server. Access to the file is managed by the license server, possibly in conjunction with a pay management system (3). Decrypted/unwatermarked media might be delivered directly to a browser (4), or it could be decoded by the appropriate DRM-enabled software application (5). Either way, a fully licensed, digital-quality media file or stream reaches to the customer (6). Key features of an effective DRM system generally include:
Data protection, so files are not easily viewed without proper privileges (Content Protection).
Unique identification of each customer to ensure that rights are applied appropriately (Fingerprinting).
Central management of rights to allow for free distribution, anti-fraud measures, and revocation (Content Authentication and legal action)
Flexibility, so the system can be tailored to various business models (rental, ownership, and read-only (Copy Control).
Rights model is the core of any content rights managements system. A rights model is a specification of the types of rights that system can keep track of or what the system can do with those rights and the attribute of those rights such as how many times content can be used, for how long user can access the content, how many times user can copy the contents, how much money etc. Rights model of DRM systems are used to define rights to content, according to some rights model, and to enforce the granting of those rights. There are three ways to enforce content rights: 1) Legally through registration forms, license agreements, and copyright laws. 24
2) Legally with an audit trail, such as copyright notices or watermarks (identifiers embedded permanently in the content). 3) Technologically, using encryption and user authentication to protect content and only make it accessible under strictly specified conditions. Content protection and tracking are the basic building blocks of every DRM system. Most of the existing DRM systems use encryption for content protection, content usage tracking, content and user authentication, etc. which cannot provide sufficient safeguard against piracy due to its limitations. On the other hand, watermarking along with encryption can ensure content protection and usage tracking. A content protection scheme that incorporates both encryption and watermarking is not foolproof but provides sufficient protection against active adversary attacks. This is likely that most successful DRM solution in the years to come, where combine encryption and watermarking can be used for content protection and related issues. In this dissertation we intend to develop content protection techniques using both encryption and watermarking.
25
CHAPTER 3
Related Work This chapter studies the theoretical aspects of the data hiding problem. Different conceptual models of data hiding problem are explored here. These models will help to comprehend the theoretical aspects of the data hiding problem. These models can be classified into two main categories: 1) the data hiding models based on communications theory, and 2) the data hiding models based on geometrical framework. Based on embedding methods the existing data hiding schemes can be divided into two classes (as discussed in Chapter 2): 1) spread spectrum based data hiding, and 2) informed data hiding. Related work in these directions is briefly discussed in Section 3.4. The goal of this chapter is to lay the foundation for the design and analysis of the data hiding systems discussed in the later chapters.
3.1
Data hiding in Communication Framework
In the recent years several researchers in the data hiding community [5, 7, 25 – 33, 39 – 71] have use traditional communications framework to analyze the theoretical-aspects (such as data hiding capacity, error probability, and performance limits) of data hiding and watermarking. A brief overview of the classical model of a communications system would be helpful to understand the similarities and differences between a conventional communications system and a data hiding system.
3.1.1 Classical Model of Communications system The channel encoder, channel decoder, and communication channel are three basic elements of the traditional communications model as illustrated in figure 3.1. Here message, m, is to be transmitted across the communications a channel.
26
Transmitter
Receiver OUTPUT INFORMATION
INPUT INFORMATION
r
x
SEQUENCE
∑
CHANNEL ENCODER
SEQUENCE CHANNEL DECODER
mˆ
m n
CHANNEL DISTORTIONS/
NOISE
Figure 3.1: Standard Communication Model
The channel encoder is a function that maps each possible message mi to a code word x, selected from a set of signals suitable for transmission over the channel. For digital communication channel encoder is generally divided into source encoder and modulator. The source encoder maps a message into sequence of binary symbols, where as, modulator maps a sequence of binary symbols into a physical signal x, suitable for transmission over the channel. In general channel encoder output is dependent on the transmission channel, but for our case x is a finite precision real sequence of length N i.e. x = {x0, x1,…, xN-1}. We also assume that these signals are bounded, i.e. these signals are power constraint, that is,
∑ ( x [i ])
2
≤ p;
p 0 , then data embedding distortion in 36
this case σ e 2 = E{(α meb) 2 } = α 2 , here b is one bit information to be embedded into host data and b = ±1. It can be shown [115, 147] that the probability of error Pe at the receiver using ML detector in the absence of noise or adversary attack is given as, Pe =
1 − e 2
2
λ
3 .7
where λ = σ e α . Hence this is clear from Eq. 3.7 that even in the absence of channel noise or attack Pe is not zero i.e. zero probability of decoding error is not attainable; this fact ( Pe ≠ 0 ) holds for other probabilistic models of the host data i.e. for gaussian or generalized gaussian host model. Therefore, additive embedding class of data hiding is not provably robust, and this is due to the host signal interference. Informed detector improves the performance of this class of data hiding methods considerably; this is because the presence of host signal at detector can be used to cancel the host signal interference. Moreover, variation of this class that minimizes the effect of host signal interference can also improve the performance. One of the most important advantages of the additive embedding based data hiding methods is that for a power-constrained transmission channel this is extremely difficult to severely degrade the host signal’s underlying pfd. Because the statistical properties of the host signal are relative invariant, this will cause a noticeable degradation of robustness performance in the presence of attack. This is because the decoder optimization criterion depends on the statistical characterization of the host signal. Therefore, the Pe does not degrade abruptly from attack free scenario to growing channel distortion. Additive embedding method such as spread spectrum watermarking is one of the first methods used for data embedding [59, 164] and still most popular one due to its advantages to withstand
37
against sever distortion and simplicity. Many variations of this method are possible depending one the nature of the host signal and the application of interest [67 – 68, 139, 140, 162, 163]. Malik et al [68] proposed frequency selective spread spectrum watermarking scheme for digital audio, in this method we use only a selected frequency range of the host signal (audio signal) for data embedding instead of the complete frequency range of the host media. Thus the host signal interference at the detector due to the selected subband signal is minimized which in return improves the robustness. The proposed method introduces low embedding distortion as watermark is embedded in the selected frequency range. Moreover, this method is capable of embedding 5 – 8 times more data compare to the existing data hiding methods of this class. Detailed overview of this method is provided in Chapter 4.
3.4.2
Informed Embedding
Recently Chen et al [30, 81] and Cox et al [71] have explored the idea of informed embedding. Data hiding methods under the informed embedding umbrella generally have higher data rate, better robustness and fidelity for bounded perturbation attack channels. These methods are capable to achieve zero-error probability as long as channel distortion is below a certain threshold [5, 6, 27, 28, and 115]. In general this class of data embedding methods use blind detector i.e. the detector has no information about the host signal Do for detection process but the encoder exploit information about the host signal to reduce the host signal interference. Cox et al’s work [71] is based on the general concept of Shannon’s paper “Channels with Side Information at the Transmitter” [34]. Where as Chen et al’s [30, 80 – 86] work is based on Costa’s work “Writing on Dirty Papers” [37]. Costa considered communication with side information at encoder over an AWGN channel as described in Figure 3.6
38
3.4.2.1 Costa’s Work The main requirement of Costa’s solution to the communication problem described in Figure 5.1 is to design an Nd –dimensional code book u N d and an appropriate encoding process; here Nd is the cardinality of the host data vector. In the limiting case i.e. as N d → ∞ Costa’s codebook u N d achieves the capacity of communication with IID Gaussian side information Do at the encoder and AWGN channel. Costa’s codebook can be defined as,
u
N
= { u [ i ] = m e [ i ] + η D o [ i ] | i ∈ {1 , 2 , . . . , N u } }
d
3 .8
where Nu is the cardinality of the codebook and η : 0 ≤ η ≤ 1 is a codebook parameter. Moreover, me ∼ N (0, σ e2 I Nd ) Do ∼ N (0, σ d2 I Nd ) n ∼ N (0, σ n2 I N d )
are the realizations of embedding pattern, host data, and channel noise which are Nd – dimensional
mutually
independent
random
processes
with
zero-mean
and
σ e2 I N , σ d2 I N , and σ n2 I N covariance matrices respectively with Gaussian pdf , where I N is and Nd d
d
d
d
–dimensional identity matrix. Costa showed [37] that the capacity of such communication system is independent of the host signal interference, that is,
C
A W G N
=
1 lo g 2
2
(1 +
σ σ
2 e 2 n
)
3 .9
which is equal to the capacity of additive spread spectrum system with informed detector [5]. 3.4.2.1.1
Quantization Index Modulation (QIM)
Costa’s scheme is purely theoretical, therefore, several practical approaches to implement Costa’s scheme have been proposed [28, 30, 57, 72 – 74, 84]. In Chen et al’s [80 – 86] proposed data hiding scheme, the host signal Do is quantized depending on the information to be 39
embedded, this scheme is commonly referred as “quantization index modulation” (QIM). For analysis and implementation, Chen et al [80 – 86] gave a low complexity practical implementation of their theoretical QIM scheme, i.e. binary dither modulation (BDM). The QIM and its variations belong to the informed embedding class of data hiding [77 – 89]. The QIM information embedding process involves modulating the index or sequence of indices with the embedding information and then quantizing the host signal with the associated quantizer or sequence of quantizers. A quantizer is approximately an identity function i.e. q ( x) ≈ x and can be uniquely described by a set Q of reconstruction points in Nd –dimensional space along with a rule of mapping the input signal of length Nd to a point in the set Q. Minimum distance rule is generally used for selecting a suitable point from Q for an input signal, therefore, different quantizers can be characterized by their reconstruction points Q only. As QIM scheme belongs to the host interference rejecting class of data embedding, therefore QIM schemes offer high data rate for power constraint attack channels [30]. Basic steps of QIM scheme can be outlined as: 1) A set of different quantizers {Q1, Q2, …, QM} is defined, where M is the cardinality of the possible embedding messages set M. 2) To embed message m, the host signal is quantized using quantizer Qm. 3) The detector quantizes the received signal Dmn using the set of all quantizors {Q1, Q2,…, QM}. Then the detector determines the index of the quantizer with reconstruction point closet to the received signal; this estimated index corresponds to the received message mˆ . 3.4.2.1.1
Binary Dither Modulation
Binary dither modulation is a low complexity implementation of the QIM in which the ensemble of the embedding functions is dither quantizers [3]. For these dither quantizers the quantization cells and reconstruction points of any given quantizer are the sifted version of the quantization 40
cells and reconstruction points of any other quantizer in the ensemble. The shifts are generally achieved using a pseudorandom vector called as dither vector, d, for information embedding purpose this dither vector is modulated according to the embedding message m. Let the modulate dither vector corresponding to the message m is denoted by d(m). The embedding function based for dither modulation is defined as [30, 81]:
D m ( D o , m ) = q ( D o + d ( m )) − d ( m )
3.10
here q (i) is a uniform scalar quantizer with step size ∆ . For binary dither modulation, the mapping from the range of the host signal values Do[n] onto the watermarked signal values Dm[n] using uniform scalar quantizer with step size ∆ is illustrated in Figure 3.7. Here, the set Q 1 (circles ‘O’) is defined by a uniform scalar quantizer with step size ∆ . Similarly, the set Q
2
(crosses ‘X’) is another uniform scalar quantizer with
same step size but with ∆ /2 offset. Do [n]
Dm [n]
O
X O ∆
X O
∆/4 ∆/4
X O
X Figure 3.7: Binary Dithered Modulation Scheme Based on Dithered Uniform Scalar Quantization
The dither vector construction and the zero-error watermark detection condition are derived as: We assume that the data embedding rate Rm is1/ N d ≤ Rm ≤ 1 , and m = {b1, b2,…, bN d Rm } is the binary representation of embedding message m where bi ∈ {0,1} for i : 1,2, …, NdRm. If ku/kc is 41
the rate of an error correcting code used for channel encoding then channel encoded binary representation of m is {z1, z2, …, z Nd / L } where L =
1 ( ku / k c ) Rm
Now two dither subvectors of length L are constructed as,
⎧ d i (1) + ∆ / 2 , d i (2) = ⎨ ⎩ d i (1) − ∆ / 2 ,
if d i (1) < 0 if d i (1) ≥ 0
i = 1, 2 , ..., L
3 .1 1
Eq. 3.11 ensures that two L –dimensional dither quantizers are at maximum possible distance from each other. Here, one dither subvector (say d(1)) is associated with binary information ‘0’ where as second dither subvector is d(2), associated with binary information ‘1’. Finally Nd/L dither subvectors associated with channel encoded bits z1, z2,…, zNd/L are concatenated to from dither vector d (m ) ∈ ℜ Nd . Finally minimum distance between reconstruction points of two quantizers Q
1
and Q
2
can be
shown [30, 84] as,
d
=
2 m in
d R
H m
k k
u c
(
∆ 2 ) 4
3 .1 2
where dH is the minimum hamming distance, a feature of the error correcting code used for channel encoding. For very small quantization cells, the mean squared distortion introduce per dimension due embedding by the uniform, scalar quantizer with step size ∆ is: E (D
E
(D
o
, D
m
∆ 2 )) = 1 2
3 .1 3
Now for bounded distortion channels and minimum distance decoding zero-error decoding condition can be shown as [82], 3d
H
k
E (D E (D o , D 4 k c N d R m σ n2 u
m
))
> 1
3 .1 4
Eq. 3.14 shows that for a fixed rate Rm, for more channel distortion energy σ n2 we need more embedding mean squared distortion i.e. E ( DE ( Do , Dm )) . Moreover, for a fixed rate and channel 42
distortion energy, then the Eq. 3.14 gives minimum perceptual distortion introduced due to data embedding. Therefore, QIM scheme gives a trade off between rate, robustness and fidelity of the data embedding process. As informed embedding class of data embedding methods use blind detector, therefore, detector can be treated as deterministic hence their performance limited by the bounded power channels distortion. For example if channel distortion σ n2 > ∆ /2 then performance of the data embedding system deteriorates and zero-error decoding is not guaranteed. We will discuss our contributions in informed embedding for digital audio data in Chapter 5. Our proposed data hiding schemes [75, 76] using phase alteration of audio data are capable to embed more data than the existing schemes while keeping embedding distortion below masking threshold. Considering the level of research activity related to data hiding in the past decade, it is evident that there has been a significant improvement in the design of data embedding and detection schemes, but at the same time sophistication in the attacks against data hiding has shown similar improvements. These parallel improvements have motivated theoretical analysis of performance limits of digital data hiding techniques. First work in this direction is by Moulin et al [39, 40, 47 – 54] where they consider digital watermarking as a game between watermark embedder and active attacker. The watermark embedder attempts to maximize the amount of embedded information whereas attacker attempts to minimize it. Other theoretical work considers the robustness against estimation attacks [5] or influence of quantization on correlation based watermark detection.
43
CHAPTER 4
Blind Data Embedding This chapter presents our initial contributions in the additive embedding or blind embedding techniques and outlines the future work. In general, additive embedding or spread spectrum based watermarking techniques embed information by adding pseudo-random sequence into the host data and correlation based detector is used for watermark detection.
4.1
Robust and High Capacity Data Embedding: Our Work
Currently we are working on the design of high capacity, robust data hiding algorithms using spread spectrum theory. Initial results of our proposed algorithm in this direction are promising [68]. Main features and performance analysis of our work are given next.
4.1.1 Data Hiding Using Frequency Selective Based Spread Spectrum As pointed out in the previous Section that the host signal interference at the detector limits the performance of additive embedding class of data hiding, therefore, we can improve the performance of these methods by either rejecting or minimizing the host signal interference. For complete rejection of the host signal interference we need informed detector which is not feasible for many data hiding applications such as copy control, device control, etc. So we can think of minimizing the host signal interference, and one possible way to do this by embedding data in a selected subband signal of the host signal instead of whole frequency band of the host signal which is the main idea of our work [68]. This frequency selective based embedding will also reduce the embedding distortion that ultimately improves the fidelity of the data hiding scheme. The frequency selective based data hiding algorithm is outlined next. 44
This method [68] is designed to overcome common shortcomings of existing DSSS based audio data hiding /watermarking systems [59 – 67] such as vulnerability to desynchronization attacks, poor detection performance, poor fidelity (inaudibility), and limited embedding capacity. Robustness to desynchronization attacks and reliability of detection performance are improved using content-adaptive features called salient points [65] of the input audio. These salient points are frame level features of the input audio signal that are invariant to common audio processing operations. Only a small fraction of the audible frequency range is used for data embedding in order to reduce the amount of audible distortion. The method exploits the frequency masking characteristics of the human auditory system (HAS) and inserts the mark into a randomly selected frequency band of the input audio signal. A secret key is used for randomly selecting a frequency band for watermark embedding. The proposed watermarking scheme induces low perceptual as well as mean squared distortion; and is therefore, the proposed scheme has high embedding capacity P. Moulin et al [39 – 41]. The detection performance of the system was investigated for a variety of signal manipulations and attacks on a watermarked audio clip. These attacks include addition of noise, resampling, requantization, filtering, and random chopping. Results show the robustness of the method, with a low detection error rate and a low bit error rate. Moreover, the proposed watermarking scheme is capable of embedding multiple watermarks in the unused frequency bands with the use of separate secret keys.
4.1.1.1 WATERMARKING USING PERCEPTUAL AUDITORY MODEL The basic idea underlying perception-based watermarking schemes is to incorporate the watermark into the perceptually insignificant region of an audio signal in order to ensure transparency. The perceptually insignificant region is determined using the human perceptual auditory model. Extensive work was done over the years on understanding the properties of HAS and applying this knowledge to audio applications [2]. An important application of perceptual 45
models is in the area of perception-based compression [165]. An important characteristic of HAS is auditory masking that has been that has been exploited in audio coding for lossy compression. We consider its use in watermarking. Human ear performs frequency analysis that maps a frequency to a location along the basilar membrane. The HAS is generally modeled as a non-uniform bandpass filter bank with logarithmically widening bandwidth for higher frequencies [165]. The bandwidth of each bandpass filter is set according to the critical band, which is defined as “the bandwidth in which subjective response changes abruptly” [2]. The critical band rate (CBR) is a measure of location on the basilar membrane just as the frequency gives a measure of location in a spectrum. The unit of critical band rate is Bark. The mapping between CBR and frequency is defined as:
(
z = 1 3 arctan (0 .6 7 f ) + 3 .5 arctan ( f / 7 .5) 2
)
4 .1
where z is CBR in Barks and f is frequency in kHz. Masking is a fundamental property of HAS and is a basic element of perceptual audio coding systems. It is a phenomenon by which a stronger audible signal makes a weaker audible signal inaudible [2], and this occurs both in frequency as well as time domain [2].
4.1.1.2
SALIENT POINT EXTRACTION
Spread spectrum techniques have been applied in digital watermarking [59 – 67] due to their potential for high fidelity, high capacity, robustness, and security. In the proposed scheme, the process of generating a watermark and embedding it into an audio signal is treated in the framework of spread spectrum theory. The original audio signal is treated as noise whereas the message information used to generate a watermark sequence is considered as data. The spreading sequence, also called pseudo-random noise sequence or PN-sequence, is treated as key. This watermarking strategy can be treated in the framework of communication models discussed in [7].
46
A critical aspect of designing a spread spectrum system is ensuring fast and reliable synchronization at the detector. Synchronization impacts performance as it reduces the overall capacity of the watermarking system, and an active adversary can use explicit synchronization information for de-synchronization attacks. To overcome these problems, synchronization is tied to attack-sensitive locations or salient points for watermark embedding and detection. Salient points are extracted based on the audio features sensitive to the HAS [65], e.g. fast energy transition points, zero crossing rate and spectral flatness measure. If these features are altered then noticeable distortion is introduced. A good salient point extraction method is one that approximately extracts the same salient points before and after common signal manipulations or watermark embedding [65]. Fast energy transition audio feature is used in our method for salient point extraction. For an audio signal Do(n): n = 0,1,2,…N-1, the short time energy ratio at each point is calculated as: E r (n ) =
E a fte r ( n ) E b e fo r e ( n )
4 .2
where Eafter(n) and Ebeforer(n) are defined as:
E b e fo r e ( n ) = E a fte r ( n ) =
∑
∑
−1 i= − r r −1
i=0
D o2 ( n + i )
D o2 ( n + i )
4 .3 4 .4
Here r is the number of samples before and after x(n). A high energy transition points are defined as: If : Er(n) > Th1 & Eafter(n) > Th2 Finally a salient point is decided as follow: 1: If two high energy transition points are separated by less than Th3 then samples are merged together to form a group. 2: Within each group, the strongest transition point is marked as a salient point. 47
here Th1, Th2 and Th3 are thresholds, these thresholds are set adaptively to ensure 3 - 4 salient points per second.
4.1.1.3
WATERMARK EMBEDDING
To generate and embed a watermark, the host data (audio) is analyzed first to determine salient points list. A block of P samples around each salient point is selected. The block is applied to a llevel modified wavelet analysis filter bank to generate (2xl-1) –subband signals of unequal bandwidths, as illustrated in Figure 4.1. Sb 9 f = fs/4 ~ fs/2
h_hp(n) X(n)
h_hp(n)
f = 0~fs/2
h_lp(n) h_lp(n)
h_hp(n)
Sb 8 f = 3fs/16~fs/4
h_lp(n)
Sb 7 f = fs/8~3fs/16
h_hp(n)
h_hp(n)
Sb 6 f = 3fs/32~fs/8
h_lp(n)
Sb 5 f=fs/16~5fs/32
h_hp(n)
:Represents downsampling by the factor of 2
h_lp(n) h_lp(n)
h_hp(n)
h_hp(n)
Sb 4 f=3fs/64~fs/16
h_lp(n)
Sb 3 f=fs/32~3fs/64
h_hp(n)
Sb 2 f = fs/64~fs/64
h_lp(n)
Sb 1 f = 0~fs/64
:Represents low pass filtering with cutoff freq. pi/2
:Represents high pass filtering with cutoff freq. pi/2
h_lp(n)
Figure 4.1: 5 –Level Modified Discrete Wavelet Analysis Filter Bank
The choice of the number of subbands is made based on a compromise between allowing a large choice in random selection and ensuring that the subband bandwidth covers at least three critical bands. A subband from lower eight bands (for l = 5) is selected using the three bit sub-key k1i, for ith salient point, where as the complete secret key K1 for subband selection is given as: K1 = k11| k12|…| k1i|…k1M where M is the cardinality of the salient point set. The selected subband is used to estimate the masking threshold Tm(k), which is calculated as follows:
48
Let sbi,j(n) for n = 0,1,2…L-1 be the jth subband of ith frame of the audio data that is selected using key K1i. Its power spectrum is defined as, Psb(k) = |Sbi,j(k)|2
4.5
here Sbi,j(k) is the discrete fourier transform (DFT) of the sbi,j(n). Now k is wrapped onto Bark scale using Eq 1. The energy in each critical band is calculated as,
E (z) =
∑
UB k = LBZ
P s b ( k ) / P z : f o r z = 1, 2 , ... z t
4 .6
where zt is the total number of critical bands in the selected subband, LB and UB are the lower and upper boundaries of the a critical band, and Pz is the total number of points in each critical band. The energy per critical band is used to calculate the masking threshold Tm(z) using MPEG layer III psychoacoustic model 1 [165]. 4.1.1.3.1
Watermark Generation
For each salient point a watermark W of length L is generated. To generate a watermark W, binary message m is mapped onto mˆ using a channel encoder. The channel encoded data is applied to binary phase shift keying (BPSK) modulator. The output of the BPSK modulator is Wm(n) : n = 0,1…q-1, where q = L/(spreading factor). Maximum length PN-sequence p of length (L/q) using log2 (L/q) bit secret key K2 is generated. Finally modulated signal Wm is spread using PN-sequence p to generate final watermark W. System key K = K1|K2. 4.1.1.3.2
Watermark Embedding
Spectral shaping based on Tm(k) of W is required to ensure inaudibility of the embedded watermark. For this purpose W(k) (DFT) and power spectrum Pw(k) of W is calculated. Now using Tm(z) inaudible DFT coefficients of the selected subband sbi,j are removed, i.e. ⎧ s bi, j ( k ) Sbn (k ) = ⎨ ⎩0
if if
P sb (k ) ≥ T m ( z ) P sb (k ) < T m ( z)
4 .7
similarly unwanted DFT coefficients of W(k) are also removed, i.e.
49
if P s b ( k ) ≥ T m ( z ) if P s b ( k ) < T m ( z )
⎧0 W n (k ) = ⎨ ⎩ w (k )
4 .8
The final watermark before embedding is given by Wf(k)=Fz•Wn(k)
4.9
where Fz is the shaping factor and defined as, F
z
=
A T m ( k ) m a x ( | W n ( k ) |)
4 .1 0
where 0 < A m
r
(z)
if
P sr (k ) ≤ m
r
(z)
4 .1 3
where Sr(k) is the DFT of sr(n) the selected subband of the received audio and Psr(k) is the corresponding power spectrum. The residual is transformed into time domain for watermark detection/extraction using IDFT i.e. rr(n)=IDFT(Rr(k))
4.14
The residual rr(n) is now used for watermark detection, by using normalized correlation test. The normalized correlation between real sequences rr(n) and PN-sequence p(n) at the detector generated using key K2 is defined as,
c o rn ( n ) =
∑ ∑
M l=− M M l=0
rr ( l ) p ( n + l )
rr ( l ) 2 • ∑
M l=0
p (l ) 2
4 .1 5
51
where L is the length of the residual signal. High correlation implies the presence of watermark as illustrated in Figure 4.3. Normalized Correlation :Watermark Present
Normalized Correlation : Watermark Absent
1
Normalized Correlation
NormalizedCorrelation
0.05 0.8 0.6 0.4 0.2
0.04 0.03 0.02 0.01 0 -0.01
0
-0.02 500
1000
1500
2000
500
1000
1500
2000
Figure 4.3: Normalized Correlation for watermarked subband (left) and unwatermarked subband (right).
The normalized correlation is compared with a threshold to determine the presence of a watermark. Let hypothesis H1 denote the presence of a watermark in a selected subband and H0 denote the absence of a watermark. The decision criterion is H
1
:
i f m a x ( c o rn ) ≥ T h 7 :
w a te r m a r k p r e s e n t
H
o
:
i f m a x ( c o rn ) < T h 7 :
w a te r m a r k a b s e n t
4 .1 6
If H1 is true then the embedded information is recovered by despreading rr(n) using the PNsequence generated using same key K2, then demodulating the resulting sequence using BPSK demodulator followed by channel decoding. The detection process is illustrated in Figure 4.4.
watermarked Audio Dm
Sb1(l) ATTACK-SENSITIVE REGION EXTRACTION
SUBBAND ANALYSIS
x(l, i)
(USING
ANALYSIS
SUBBAND
MODIFIED
WAVELET ANALYSIS
l = 1~M
FILTER)
sp i=1~L AUDIO CONTENT
Key : k1i
Sbp(l)
SELECTION
Key : k2
OUTPUT WATERMARK CHANNEL
PN-SEQUENCE
Sbj(l)
DECODING
GENERATOR
MASKING THRESHOLD
peak ≥ Th
EXTRACTION PSYCHOACOUSTIC
(USING
MODEL)
WATERMARK
BPSK
SPREADING
DEMODULATION
SEQUENCE
100101101.. .
CORRELATOR
Tma (z)
DETECTOR
RESIDUAL EXTRACTION
(USING
MASKING
peak < Th
THRESHOLD)
No watermark
Figure 4.4: Block Diagram for Watermark Detection
4.1.1.5 EXPERIMENTAL RESULTS 52
The robustness of the proposed scheme was tested on speech signals and music. The tests included several degradations and distortions, i.e. addition of noise, lossy compression, low pass filtering, resampling, random chopping, and multiple watermarks. The detection performance in each case depends on the following measures, 1) watermark detection rate (WDR) which is a measure of watermark detection, and 2) the bit accuracy rate (BAR) which is a measure of data recovery. The bit accuracy rate is defined as, BAR =
N u m b e r o f B its C o r r e c tly D e te c te d N u m b e r o f B its E m b e d d e d
4 .1 7
and watermark detection rate:
WDR =
Number of Watermarked Frames Correctly Detected Number of Watermarked Frames Embedded
4.18
The overall performance of the system is defined as,
DPM = BAR ×W DR
4 .1 9
where DPM stands for detection performance measure. Detection results for degraded watermarked audio based on DPM for a variety of conditions are described below.
White Gaussian noise is added to the watermarked audio; the DPM (as defined in Eq. 19) values in the presence of white gaussian noise with power from 0 to 50% of the signal power are shown in Figure 4.5.
53
Detection Performance Measure vs Noise Power 1.005
1
Detection Performance Measure
Noise Power = percentage of the Audio Power
0.995
0.99
0.985
0.98
0.975
0
5
10
15
20
25 Noise Power
30
35
40
45
50
Figure 4.5: DPM for different values of noise power (Pn)
Watermarked audio is down-sampled to 22.05 kHz and then interpolated to 44.1 kHz. The DPM value for this test is 1.
Watermarked audio undergoes ISO/MPEG-1 Audio Layer III encoding/decoding [165] at a bit rate of 128 kbs. The DPM value for compression test is 1.
Watermarked audio signal is lowpass filtered with 4 kHz cutoff frequency, Detection of resulting audio gives a DPM of .995. Detection performance is still acceptable despite severe audible distortion
To investigate desynchronization attacks, one out of every 100 samples of watermarked signal was randomly dropped. Detection applied to this signal gave a DPM of 1.
Three watermarks simultaneously embedded in the audio, with a unique sectary key assigned to and a unique subband selected for each watermark. The DPM was 1 as long as the number of watermarks is less than the number of analysis subbands.
54
4.2
Future Directions
A novel watermarking scheme for audio based on FS-DSSS is proposed. The technique introduces low mean squared as well as perceptual distortion compare to existing spread spectrum schemes [59 – 67] this is due to that fact that a watermark is embedded in a small frequency band of complete audible frequency range. The watermarking capacity theory presented in [39 – 41] suggests that the proposed scheme can embed more information. The proposed method is also robust to standard data manipulations i.e. noise addition, compression, random chopping and re-sampling.
5.2.1
Proposed Dada Hiding Scheme for Images
We are currently investigating to extend our frequency selective watermarking scheme [68] for digital image watermarking, image authentication, and image fingerprinting. As the existing image watermarking schemes [5 – 24] generally use blind detector for watermark detection which limits the performance of blind or additive data embedding schemes due to host signal interference. This interference can be reduced if watermark is selected from the null space of the host data. This will improve the detection performance of the data hiding scheme because watermark selected from null space of the host signal therefore the host signal interference is minimum. Bayesian source separation approach can be used for blind watermark detection. As embedded watermark is orthogonal therefore separation matrix estimation would be computationally efficient as well. For image decomposition we are investigating to use l –level discrete wavelet analysis filter bank. The reason for using DWT for watermark embedding is multi-folds: 1) the wavelet transformation provides good space-frequency localization to analyze image features such as, edges or texture areas, 2) due to multiresolution representation of the image, hierarchical processing of the image is possible that can be used for progressive watermark 55
embedding/decoding, 3) the wavelet transform is very flexible to adapt a give set of images or application at hand, 4) wavelet coefficients can be generally characterized by a Gaussian distribution [166, 167 ] which will improve the computationally efficiency of separation matrix estimation using Bayesian approach, and 5) wavelet transform is compatible to JPEG2000, the most recent still image compression standard. For watermark embedding all subbands except l th –level approximation subband are selected because embedding in this subband will degrade the visual quality of the watermarked image. Moreover, as diagonal subbands are generally less sensitive to the quantization noise therefore, these subbands at 1st and 2nd level are not suitable for watermark embedding, because these subbands are generally discarded during lossy compression. The subbands along horizontal and vertical orientation at 2nd and 3rd level are suitable for watermark embedding. Figure 4.6 illustrates 3 –level wavelet decomposition. LL4
HL4
LH4
HH4
HL3 HL2
LH3
HH3 HL1
LH2
HH2
LH1
HH1
Figure 4.6 Three Level Wavelet Decomposition
The proposed frequency selective image watermarking using DWT scheme is outline as:
Decompose the image I using l –level discrete wavelet analysis filter bank.
Randomly select a subband using secret key from the decomposed image such that
{
}
x j = sbiθ : i ∈ 1, 2,3...l ∀θ ∈ 2, 4 where j = 1,2.
56
Generate a watermark w with non-Gaussian distribution using secret key such that watermark lie in the null space of the selected subband.
Calculate the masking threshold for the selected subband using noise visibility function [167].
Embed watermark in the selected subband as: xe = x + α xsw
4 .2 0
where xe is the watermark embedded subband, xs are the wavelet coefficients above the masking threshold, and α is the level adaptive scaling factor.
Reconstruct the watermarked image using discrete wavelet synthesis filter bank.
Now to detect the watermark using blind detector some variations of blind source separation scheme can be used but we are thinking to use Bayesian source separation framework. Because the embedded watermark is uncorrelated to the subband in which it is embedded therefore separation matrix estimation is possible. For image authentication and fingerprinting applications, original image is available at the detector therefore detection performance will definitely improve. The proposed scheme is flexible enough that we can modify it depending on the application of interest. For example, for image fingerprinting application, to detect the fingerprint we can use normalized correlation function because the host data is available at the detector. Moreover, multiple fingerprints can be embedded simultaneously.
4.2.2 Proposed Dada Hiding Scheme for Video Video and audio data types need more attention of the researchers from the data hiding community because they are strong candidates to carry more data than digital images. Moreover, researchers explored these data types very little in the past. Entertainment industry is facing more losses due to digital audio and video instead of the digital images. Audio and video data hiding 57
can be used for multimedia wireless communication to improve the QoS, enhanced performance for intended recipients where as normal service for a regular user. Image watermarking technique proposed in the previous section can be extended for video data hiding with certain modifications. For video data hiding I –frames and motion compensation vectors can be used for data hiding where as P –frames can be used as a backup or 2nd level data hiding. To eliminate the frame replacement or frame shuffling attacks in video data hiding, frame pairing can be used for data embedding i.e. information about frame f is embedded in frame f’ where as index number of f’ is greater than the index number of f. We are also investigating to develop data hiding scheme for efficient error concealment of multimedia transmission over lossy and busty channels.
Chapter Summery This chapter gives an overview of the blind embedding schemes. In Section 4.1 we discus the existing additive data hiding schemes, their advantages, and their limitations. Our contribution in this class of data hiding schemes is provided in Section 4.2. Low host signal interference, reduced embedding distortion, high data rate, and robustness against adversary attacks are the attractive features of the proposed scheme. The simulation results show the robustness of the proposed scheme against common data manipulation attacks. Possible extension of our frequency selective watermark embedding scheme to digital images and video are proposed in section 4.3. For watermark detection we are planning to use Bayesian source separation approach. We are also investigating to use data hiding for error concealment of multimedia transmission over a busty channel.
58
CHAPTER 5
Informed Data Embedding This chapter provides our initial contributions in the informed embedding or host interference rejection based embedding techniques and outlines the future work. In general, informed embedding based watermarking techniques embed information using informed encoder i.e. by exploiting the host signal knowledge at encoder and use blind detector for watermark extraction.
5.1
High Rate Data Embedding Using Informed Encoding: Our Work
This is clear from Eq. 3.14 that QIM based data embedding techniques introduces large embedding distortion for more robust data embedding at fixed data rate. Moreover, QIM schemes did not pay attention to the human perceptual system. Therefore, for low perceptual distortion and better robustness performance, data embedding schemes have to incorporate the human perceptual system. Malik et al [75, 76] proposed a data hiding using deterministic dithering in the selected frequency range of the audio signal for data embedding. The frequency range selection for dithering is based on the human perceptual model.
5.1.1 Data Hiding Using Frequency Selective Dithering (Our Contribution) In [75, 76] we propose a novel perception-based high capacity data hiding methods. In this scheme we explore the following properties of HAS: the magnitude distortion at a specific frequency in an audio signal is inaudible if it is below masking threshold; and human perception is less sensitive to absolute phase changes in a certain frequency range [2, 154]. Not all of the customary full range of audible frequencies, i.e. 20 Hz ~ 20 kHz, is suitable for data embedding. In the higher frequency range (≈ f > 10 kHz) detection of small magnitude changes is unreliable due to insignificant signal energy. On the other hand human perception is more sensitive to phase distortion in the lower frequency range (≈ f