ADVANCED TECHNIQUES IN DIGITAL WATERMARKING AND DATA HIDING

Transcription

1 Hellenic Open University School of Science and Technology Digital Systems & Media Computing Laboratory ADVANCED TECHNIQUES IN DIGITAL WATERMARKING AND DATA HIDING Eleftherios K. Chrysochos June 2009, Patras

2 ii

3 Dedicated to my friends and family iii

4 iv

5 Abstract This thesis is entitled Advanced Techniques in Digital Watermarking and Data Hiding. Two close but distinct research areas are related to this thesis. In both areas the main goal is to embed information in a digital medium. The main difference between them is the scope of the application. In data hiding the medium-carrier is of no importance, while in digital watermarking the medium is of crucial importance as well as the message hidden within. Therefore research in both areas sometimes overlaps and certain embedding techniques are shared in both fields. The main research interest is digital robust watermarking. A reversible, blind, watermarking scheme which is based on histogram modification is presented. The watermark is robust against geometrical attacks but susceptible to compression and filtering attacks. To cope with this kind of attacks our research turned to frequency domain and chaos. A watermarking scheme which is v

6 based on a chaotic function for embedding and a correlation method for detection was introduced. This novel scheme shows increased robustness against JPEG compression and filtering attacks. Research in reversible embedding and especially in Difference Expansion (DE) techniques led to a novel reversible DE transform which embeds two bits of information in a triplet of coefficients. This transform can be applied in spatial or frequency domain and induces minimum distortion to the initial coefficients of the image. Another implementation in image data hiding domain was a scheme based on DE that uses consecutive, overlapping pairs, instead of the non-overlapping pairs or triads (used by traditional DE techniques) outperforming all existing DE schemes, both in terms of capacity and PSNR. vi

7 Declaration This is to certify that: (i) the thesis comprises only my original work towards the PhD except where indicated, (ii) due acknowledgement has been made in the text to all other material used, (iii) the thesis is less than 100,000 words in length, exclusive of table, maps, bibliographies, appendices and footnotes. Eleftherios K. Chrysochos vii

8 viii

9 Publications Chrysochos E., Fotopoulos V., Skodras A., Xenos M., "Reversible Image Watermarking Based on Histogram Modification", 11th Panhellenic Conference on Informatics with international participation (PCI 2007), Vol. B, pp , Patras, Greece, May, Chrysochos E., Fotopoulos V., Skodras A., "Robust Watermarking of Digital Images Based on Chaotic Mapping and DCT", 16th European Signal Processing Conference (EUSIPCO 2008), Lausanne, Switzerland, August, Chrysochos E., Fotopoulos V., Xenos M., Stork M., Skodras A. N., Hrusak J., Chaotic-Correlation Based Watermarking scheme for Still Images, IEEE International Conference: Applied Electronics 2008, Pilsen, Czech Republic, September 10-11, Chrysochos E., Varsaki E., Fotopoulos V., Skodras A., " High Capacity Reversible Data Hiding using Overlapping Difference Expansion", 10th International Workshop on Image Analysis for Multimedia Services (WIAMIS 2009), London, UK, 6-8 May, ix

10 x

11 Acknowledgements This work was funded by the European Union - European Social Fund (75%), the Greek Government - Ministry of Development - General Secretariat of Research and Technology (25%) and the Private Sector in the frames of the European Competitiveness Programme (Third Community Support Framework - Measure programme ΠΕΝΕΔ - contract no.03εδ832). xi

12 xii

13 Contents 1. INTRODUCTION OVERVIEW OF DIGITAL WATERMARKING DOMAIN FOR IMAGES HISTORICAL REVIEW CRITERIONS FOR DIGITAL IMAGE WATERMARKING CATEGORIZATION VISIBILITY Visible- Invisible Watermarking IDENTIFICATION METHOD FOR THE WATERMARK Readable- Detectable Watermarks REVERSIBILITY Reversible (lossless) - Non reversible schemes ROBUSTNESS Robust Fragile Semi fragile watermarks INFORMATION REQUIRED FOR WATERMARK EXTRACTION Blind Non blind Zero knowledge - Zero watermarking NUMBER OF WATERMARKS USED Simple Multicasting ADJUSTABILITY st 2nd generation watermarking systems HUMAN VISUAL SYSTEM (HVS) HVS based Non HVS based systems WATERMARKING DOMAIN Spatial Frequency Histogram ICA Spatial Domain Frequency Domain Histogram Domain Independent Component Analysis (ICA) HYBRID TECHNIQUES REVERSIBLE WATERMARKING BASED ON HISTOGRAM MODIFICATION EXISTING SCHEMES PROPOSED WATERMARKING SCHEME Features of the Proposed Scheme Embedding Procedure Watermark Extraction Restoration of Watermarked Image EXPERIMENTAL RESULTS RESEARCH CONTRIBUTION WATERMARKING AND CHAOS CHAOS AND WATERMARKING EXISTING SCHEMES CHAOTIC MAP PROPOSED SCHEME Features of the proposed scheme Embedding Procedure Extracting Procedure EXPERIMENTAL RESULTS xiii

14 4.5 RESEARCH CONTRIBUTION ADVANCED TECHNIQUES IN DATA HIDING DIFFERENCE EXPANSION METHODS Tian s Difference Expansion Transform (baseline method) Shen et al s 3C2B Transform on Triplets Alattar s Transform on triplets NOVEL DE TRANSFORM ON TRIPLETS Computational Cost Analysis Performance Analysis with regard to Coefficient Alteration Shen et al s 3C2B Transform on Triplets Alattar s 3C2B Transform on Triplets Proposed Novel Transform Experimental Results Research Contribution PROPOSED DATA HIDING SCHEME USING OVERLAPPING DE Embedding Procedure Extracting Procedure Capacity Issues Experimental Results Research Contribution CONCLUSIONS FUTURE WORK REFERENCES xiv

15 Chapter 1 - Introduction 1. Introduction In Recent years the majority of the information that involves sound, image and video is stored in digital form. Multimedia in digital form offers many advantages and new potentials to the average user. Likely the most common used potential of digital media is the untroubled copy without degradation of the medium. Another convenience of digital multimedia is the ability of easy modification of its content. The above actions may be permissible, like the legitimate copy of a medical digital image for remote diagnosis purposes, or non permissible, like the illegal copy and distribution of a digital music album. After the 1980 s the personal computers invaded everyday life and step by step became a necessity for every home and business. Another phenomenon of the digital era is that music albums as well as films and photographs are no longer produced or distributed in analog form or hard copy. In our days all multimedia are almost exclusively distributed in digital form. The above two factors lead to global use of digital multimedia and a - 1 -

16 Chapter 1 - Introduction simultaneous rapid increase of copyright piracy as copy and modification of digital content was at hand by an average user. Another crucial factor that also contributed to the increase of piracy is the free and untroubled traffic of information over the World Wide Web. Nowadays everybody with no special skills on computer science can easily surf through the internet (World Wide Web) simply by using a web browser (e.g. internet explorer or firefox) or query and acquire information with the use of a search machine (e.g. Google). Often this information involves multimedia files. Sometimes this is a legitimate action but other times when the multimedia file is copyrighted, sharing or copying its content is considered illegal and unethical. Nevertheless sharing copyrighted material over the internet is a very common practice in our days. A system that would protect digital content was needed in order to protect and secure its copyright. In some cases a question of validity or credibility arises. This is the case when a digital content is used in sensitive applications, like medical or military applications. In these cases a system of validation and credibility check is required in order to avoid critical errors. Digital - 2 -

17 Chapter 1 - Introduction Watermarking is a developing field of research that copes with these issues. Music industry as well as movie making industry suffers each year from huge economic losses due to music and movie piracy respectively. This also affects the multimedia industry in general as all multimedia content is nowadays produced and distributed in digital form and therefore is subsequent to digital piracy. Due to these enormous economic losses involved in digital piracy there is an increasing interest in the area of digital watermarking as digital watermarking applications could be used as countermeasure against digital piracy

18 Chapter 2 Overview of Digital Watermarking Domain for Images 2. Overview of Digital Watermarking Domain for Images Digital watermarking stands for embedding a signature signal, called watermark, into a digital cover, in order to prove ownership, check authenticity or integrity of the cover, and it may relate to audio, images, video or even text. This thesis focuses on digital watermarking for images. 2.1 Historical Review The first record of visible watermarking was applied on paper in Italy in 1282 B.C. [11]. This method was introduced by papermakers in order to mark and identify their products. During the electronics revolution, the first electronic watermarking application Identification of sound and like signals, was introduced by Hembrooke E. in 1954 [17] and related to audio signals. His patent described a method for imperceptibly embedding an identification code into music for the purpose of proving ownership

19 Chapter 2 Overview of Digital Watermarking Domain for Images The first publications about digital image watermarking were seen in 1979 while the area of digital image watermarking developed rapidly during the 90 s [11]. 2.2 Criterions for Digital Image Watermarking Categorization The various existing watermarking systems can be categorized according to various factors. These factors may involve the procedure used for the embedding of the information in the image or the characteristics and the attributes of the new watermarked image [11], [4], [32] and [12]. The criterions used for categorization of watermarking schemes are: Watermark visibility: whether the watermark is visible via the naked eye. Identification method for the watermark. Reversibility: whether the watermarked image can be restored in its original status. Robustness: how robust is the watermark against - 5 -

20 Chapter 2 Overview of Digital Watermarking Domain for Images compression, geometrical attacks and filter attacks. Information required for identification and authentication of the watermark embedded in an image. Simple or multicasting: whether more than one embedding procedures may be applied for one image. Adjustability of the watermarking algorithm. Human Visual System (HVS): whether the watermarking procedure takes into account the particular features of the HVS. Watermarking domain. According to the above criterions watermarking systems are divided to different categories as shown above. 2.3 Visibility Visible- Invisible Watermarking Visible watermarking is the procedure during which a signalsignature is embedded in an image in a perceivable way usually to declare the creator or the copyrights owner of the image. Common - 6 -

21 Chapter 2 Overview of Digital Watermarking Domain for Images example of visible watermarking over video is the exclusive video clips shown occasionally on television and bear the logotype of the TV station. In visible watermarking obviously there is optical degradation of the original signal. Invisible watermarking on the other hand embeds a signature in the digital medium which is not perceivable via naked eye. Applications of visible watermarking were invented prior to invisible and electronic watermarking [6]. 2.4 Identification Method for the Watermark Readable- Detectable Watermarks In detectable watermarking systems the system itself is responsible for determining whether an image carries a watermark or not. This is accomplished by comparing a correlation factor according to a specific, predetermined or adjustable threshold [31], [49]. If the correlation factor is higher than this threshold the presence of a watermark is certified and the image is considered - 7 -

22 Chapter 2 Overview of Digital Watermarking Domain for Images watermarked. Otherwise the image is considered nonwatermarked. In watermarking systems that use readable watermarks the identification of the watermark lies on the human factor. The watermark used for embedding is a logotype or another (usually smaller) image. The logotype usually is a binary or a greyscale image. Such systems are presented in [46], [20]. 2.5 Reversibility Reversible (lossless) - Non reversible schemes A watermarking scheme is considered reversible or lossless when the original image can be fully restored after extracting the watermark of the watermarked image. Full restoration of an image means that the restored image must be equal in every way with the original (host) image without any degradation (perceivable or not). Reversible watermarking schemes may use one or more keys for extracting and restoring procedures. These that use two keys (private and public) are called asymmetric. Usually the first key is public and used for the - 8 -

23 Chapter 2 Overview of Digital Watermarking Domain for Images extracting and identification of the watermark, while the second one is private and is used for the removal of the watermark and restoration of the image back to its original status [20], [40]. On the other hand when the full restoration of the watermarked image is not possible, the watermarking scheme is considered non reversible. The lack of reversibility in these cases is subsequent of some non reversible procedures like truncation, approximation, quantization etc. which take place during the embedding procedure. 2.6 Robustness Robust Fragile Semi fragile watermarks A digital signal may sustain various alterations- attacks, malicious or not. When the watermark after sustaining a specific attack, is still detectable (not erased) then it is considered robust against the specific attack. When the watermark is robust against various attacks is called robust [24]. Robust watermarks are usually used in copyright applications

24 Chapter 2 Overview of Digital Watermarking Domain for Images Watermarking schemes which aim for robustness generally use methods for synchronizing the watermark (procedure required usually after attacks) as preprocessing before detecting the presence of the watermark [33]. When the watermark is designed on purpose not to withstand any alteration of the watermarked image it is considered fragile. Fragile watermarks are commonly used for authentication applications [49], [20], [40]. Any minor alteration of the image will destroy the watermark making it non detectable and therefore the image non authentic. Fragile watermarking is usually used for integrity check and authentication applications. However there are some watermarking schemes that exhibit robustness against some attacks (e.g. compression) but are vulnerable to other attacks (e.g. geometrical attacks or low pass filtering). These watermarking systems are called semi fragile. Most of the watermarking schemes fall into this category since it is very difficult to design and algorithm that can endure all sorts of attack. Nevertheless as new more robust algorithms are invented, new pattern of attacks are invented as well

25 Chapter 2 Overview of Digital Watermarking Domain for Images 2.7 Information Required for Watermark Extraction Blind Non blind Zero knowledge - Zero watermarking Watermarking systems are categorized according to the information required in order to trace and identify the watermark in a digital medium. Therefore a watermarking system is considered blind or oblivious when the initial (non watermarked) image is not needed for the extracting procedure [46], [25]. In case the original image (as well as the watermarked image) is necessary for the extracting procedure the watermarking scheme is called non blind or non oblivious. As expected non blind watermarking schemes show increase robustness against attacks and are base on comparing the watermarked image with the original (non-watermarked) one. On the other hand this process is not always convenient since the host image may not be at hand. Therefore blind systems are more

26 Chapter 2 Overview of Digital Watermarking Domain for Images flexible and have more applications. Blind systems are commonly used for copyright applications. Zero knowledge watermarking systems are characterized by the fact that for the extracting procedure they only require the watermarked image. In this case no additional information (e.g. a security code, a private or public key, etch) are needed for the algorithm to decide whether an image is watermarked or not and to retrieve the respective watermark [49]. Zero watermarking techniques do not embed information in the host image. They usually extract from the image some predefined characteristics that are used to form a unique signature for the specific image [31]. In this way there is no degradation in the host image

27 Chapter 2 Overview of Digital Watermarking Domain for Images 2.8 Number of Watermarks Used Simple Multicasting In simple watermarking the watermark is embedded in the host image only once [39]. In multicasting an already watermarked image can go through the embedding procedure without destroying the precedent watermark. In this way an image can be watermarked successively many times with different watermarks or the same watermark can be embedded in one image several times (aiming in increasing the robustness of the algorithm against malicious attacks) [28]. 2.9 Adjustability 1st 2nd generation watermarking systems The basic difference between first and second generation systems is the capability of adjusting the watermarking algorithm according to the features of the host image and the watermark to be embedded

28 Chapter 2 Overview of Digital Watermarking Domain for Images First generation systems do not take into account the specific features of each image and the algorithm used is standard [39], [16]. Second generation systems on the contrary adjust the embedding algorithm according to the features of the image [28], [3]. Second generation watermarking systems provide better results in terms of robustness and quality of the watermarked image as well as in terms of capacity. Nevertheless second generation algorithms exhibit increased complexity which results in higher computational cost in the embedding and extracting procedure compared to first generation systems Human Visual System (HVS) HVS based Non HVS based systems In recent years watermarking systems tend to take into account the special features of the Human Visual Systems (HVS). These watermarking schemes make sure that the alterations imposed by the watermarking procedure on the host image are not perceivable via the HVS [2]. In order to achieve this they use

29 Chapter 2 Overview of Digital Watermarking Domain for Images special quality metrics adapted to the HVS characteristics [3] instead of the standard Peak Signal to Noise Ratio (PSNR) or other Means Square Error (MSE) metrics Watermarking Domain Spatial Frequency Histogram ICA Watermarking techniques can be categorized according to the domain used to embed the watermark. Therefore there is spatial, frequency and histogram domain, which are more commonly used in watermarking and algorithms that are based on Independent Component Analysis (ICA) and Singular Value Decomposition (SVD) transforms. Each domain has certain advantages and disadvantages which will be analyzed above Spatial Domain The first attempts on digital image watermarking were in spatial domain. The main advantage of spatial watermarking is simplicity. Therefore spatial schemes have low computational

30 Chapter 2 Overview of Digital Watermarking Domain for Images complexity and consequently need less computational time. Another advantage is the high capacity offered for the watermark during the embedding procedure. The disadvantage of spatial domain is that the watermark is vulnerable to attacks (low robustness) [16]. Many spatial techniques embed the watermark information in the Least Significant Bit (LSB) of every pixel and are called LSB techniques [2]. Each LSB scheme embeds the watermark with different algorithm but they all take advantage of the insensitivity of the Human Visual System to minor changes of colour and brightness respectively. However spatial methods are extremely susceptible to any kind of attack. An interesting approach in spatial domain was introduced by Tian J. [39] who used Difference Expansion (DE) for embedding information in an image. This method is fully reversible, without imposing significant degradation ot the watermarked image. Nevertheless it is vulnerable to attacks since it is basically an LSB technique and inherits all itd disadvantaged. The main advantages of this technique are that it is blind, reversible and provides high capacity for the watermark. Specifically the capacity s upper bound is given when for each pair of pixels a watermark bit can be

31 Chapter 2 Overview of Digital Watermarking Domain for Images embedded (e.g. in an image of 640X480 pixels bits of information may be embedded) Frequency Domain Most popular watermarking systems are these that work in frequency domain. These systems make use of some transforms like Discrete Fourier Transform (DFT), Discrete Cosine Transform (DCT) and Discrete Wavelet Transform (DWT). The main advantage of Discrete Fourier Transform (DFT) is that its coefficients stay unaltered after translation attacks. DFT is a complex transform and during this transform the image is divided into two matrices, of amplitude and phase. The phase matrix is more crucial for the quality of the image. Therefore embedding the watermark in the phase matrix makes it more robust against attacks (inducing however more degradation in the quality of the image). This is also coherent with the communication theory that states that frequency modulation is more robust to noise in comparison to amplitude modulation. Watermarking schemes that use DFT transform have some disadvantages too. In order to have

32 Chapter 2 Overview of Digital Watermarking Domain for Images real values for the image luminosity or colour after the inverse DFT transform (idft), the conjugate complex factors must remain symmetric. This demand of symmetry divides in half the given space for information embedding reducing the capacity of the scheme in half. Another disadvantage is that DFT coefficients and especially phase coefficients are susceptible to compression attacks (e.g. JPEG, MPEG). Discrete Cosine Transform (DCT) gives as a result real coefficients. The basic characteristic of this transform is the high concentration of energy in low frequency coefficients with relative low computational cost [38]. This is shown in figure 1 where the DCT coefficients are in logarithmic scale and have been normalized with regard to the DC coefficient

33 Chapter 2 Overview of Digital Watermarking Domain for Images DCT Fig1. Dandelion image and the respective DCT power spectrum Most common lossy or lossless compression techniques (e.g. JPEG, MPEG1, MPEG2 and H26x) use as basis DCT transform. Another feature of this transform is that it is orthogonal and separable. That means that it can be applied on a two dimensional matrix (i.e. image) separately in each dimension thus keeping computational cost low. DCT watermarking schemes exhibit high capacity and increased robustness against attacks like compression and low pass filtering. Furthermore by using DCT watermarking in video (or image) compressed domain (JPEG, MPEG1, MPEG2, H26x) there

34 Chapter 2 Overview of Digital Watermarking Domain for Images is no need to decode and re-encode the signal, saving a significant computational time. Another interesting transform used in digital watermarking is Discrete Wavelet Transform (DWT). It is used both in image and video watermarking as it has many advantages [21], [9]. DWT is a hierarchical transform and analyzes the signal into different bands and levels. It supports resolution of a signal in successive levels as shown in figure 2 where a signal is analyzed in three levels. X[n] stands for the initial signal and h[n], g[n] stand for a high-pass and low-pass filter respectively. Fig2. 3 level DWT analysis. X[n] is the initial signal and h[n], g[n] are high and low pass filters respectively

35 Chapter 2 Overview of Digital Watermarking Domain for Images DWT is a separable transform (like DCT) and it can be applied on a two dimensional matrix (i.e. image) separately in each dimension. Two dimensional DWT (2D-DWT) analyzes the initial image in four frequency bands. The first band (LL) contains the low frequencies both in horizontal and vertical direction. The second band (HH) contains the high frequencies both in horizontal and vertical direction. The third band (HL) contains the high frequencies in horizontal direction and low frequencies in vertical direction. The forth band (LH) contains the low frequencies in horizontal direction and high frequencies in vertical direction. This is shown in figure

36 Chapter 2 Overview of Digital Watermarking Domain for Images Fig3. Analysis of Lena image in frequency bands (LL, HL, LH and HH) by DWT transform The LL band represents the approximation of the image and is the most significant band as it caries most of the image energy. The other three bands contain the details of the image (texture of the

37 Chapter 2 Overview of Digital Watermarking Domain for Images image). The Frequency bands are separated by successive filtering and down sampling of the image first in horizontal and afterwards in vertical direction as shown in figure 3. In 2004 Kundur and Hatzinakos presented Fusemark [18], a new invisible and robust image watermarking scheme which affected many following works. In this scheme DWT is applied both on the original image and the logotype to be embedded. The DWT coefficients of the image and the DWT coefficients of the logotype are then merged. This method is very robust to attacks but has the drawback that it is not blind (i.e. the initial image is needed for extracting the watermark). Gui Xie and Hong Shen evolved fusemark and created XFusemark [46], a blind watermarking scheme which is based on the special features of the Human Visual System (HVS based). Xfusemark applies DWT on the image and DCT on the watermark logotype and is based on fusion of the respective DWT and DCT coefficients. A novelty of Xfusemark is that it embeds the DCT coefficients in the LL band of the image (image approximation) in specific coefficients according to a predefined key. This was done in order to embed as much watermark energy possible in the

38 Chapter 2 Overview of Digital Watermarking Domain for Images image with respect to maintaining the image quality according to the HVS. The final watermarked image is generated by applying inverse DWT on these (fused) coefficients. Xfusemark is very robust to compression attack and Gaussian noise Histogram Domain A group of watermarking systems embed information in a signal through histogram modification. Such systems can not be considered as neither spatial nor frequency systems. One interesting watermarking system that works in histogram domain was introduced by Zhicheng Ni et al [26]. The specific system works on the luminosity histogram of an image (fig. 4) is fully reversible and provides rather high capacity for the watermark with low computational cost. The only disadvantage of this system is its low robustness against attacks. An improvement of [26] was proposed by Lee Sang-Kwang, Suh Young-Ho and Ho Yo-Sung [20]. In this scheme the capacity of the watermark is significantly increased since a difference image, generated as the difference of each pixel of even line from

39 Chapter 2 Overview of Digital Watermarking Domain for Images the respective pixel of the next odd line, is used for embedding the watermark. The disadvantage of this scheme is the same with its predecessor, that is low robustness against attacks, as they are both LSB techniques. Fig4. Luminosity histogram of Elaine image

40 Chapter 2 Overview of Digital Watermarking Domain for Images Independent Component Analysis (ICA) Another category of watermarking schemes is based on Independent Component Analysis (ICA). ICA is a signal processing and data analysis method which was firstly used in Blind Signal Separation (BSS). BSS is applied for acquiring two statistically independent signals from their mixed signal. Watermark extraction can be considered as a BSS problem (decorelating two statistical independent signals), where the overall signal is the watermarked image, and the two statistical independent signals are the host image and the watermark respectively. In recent years ICA algorithms are used in pattern recognition, image compression and analysis as well as in watermarking applications [25], [37], [35]. Watermarking systems that use ICA are usually blind and robust against attacks. Nevertheless ICA algorithms are computationally demanding making the watermarking system rather slow due to high computational time, excluding it from real time applications

41 Chapter 2 Overview of Digital Watermarking Domain for Images 2.12 Hybrid Techniques There are numerous watermarking systems that do not fall in the categories presented above either by combining characteristics by several categories, or by a applying a completely different approach, constituting consequently a category of their own. Such an approach is [21] where the watermark is embedded by modifying the histogram of the Integer Wavelet Transform (IWT) coefficients of the image. The modification of the histogram is made with respect to the properties of the Human Visual System in order to reduce the degradation of the image. Yu et al [49] use multiresolution wavelet transform and embeds a complex watermark. Specifically the real part of the watermark coefficients and the imaginary part of the watermark coefficients are embedded in different bands while keeping the magnitude of the coefficients stable. In this way the algorithm can detect the existence of a watermark with no further information than the watermarked image (zero knowledge)

42 Chapter 2 Overview of Digital Watermarking Domain for Images Another interesting work that could be characterized as hybrid was presented by Osborne et al [28]. The image is separated in two regions, Region Of Interest (ROI) and Region Of Background (ROB). In Region Of Interest a fragile watermark is embedded, while in the Region Of Background (ROB) a robust watermark is embedded many times (multicasting). The robust watermark is formed by the DCT coefficients of the ROI. In that way any alteration of the critical ROI is traced (by the fragile watermark), while an approximation of the ROI can be reconstructed by the ROI DCT coefficients included in the robust watermark Watermarking Taxonomy The overall taxonomy of the various watermarking systems is shown in figure

43 Chapter 2 Overview of Digital Watermarking Domain for Images Watermarking Systems Identification Method Information Required for Watermark Extraction Watermarking Domain Number of Watermarks Visibility Reversibility Adjustability Robustness HVS Visible Readable Blind Reversible 1st generation Robust Spatial Simple HVS based Non Visible Detectable Non blind Non reversible 2nd generation Fragile Frequency Multicasting Non HVS based Zero knowledge Semi-fragile Histogram Zero watermarking ICA Fig5. Taxonomy of Watermarking Systems Hybrid Systems

44 Chapter 3 Reversible Watermarking Based on Histogram Modification 3. Reversible Watermarking Based on Histogram Modification Some watermarking systems embed information in a signal through histogram modification. Such systems can not be considered as neither spatial nor frequency systems. 3.1 Existing schemes One interesting watermarking system that works in histogram domain was introduced by Zhicheng Ni et al [26]. The specific system works on the luminosity histogram of an image (fig. 4) is fully reversible and provides rather high capacity for the watermark with low computational cost. The only disadvantage of this system is its low robustness against attacks. Zhicheng Ni et al s algorithm initially locates the histogram bins with the maximum and minimum inputs (where input is the number of pixels with luminosity value equal to the respective bin) as shown in figure

45 Chapter 3 Reversible Watermarking Based on Histogram Modification Max Min Fig6. Maximum and minimum bins of luminosity histogram for Elaine image Then the part of the histogram bounded by these bins (maximum and minimum) is shifted to the direction of the minimum. This means that the luminosity of all pixels with luminosity values between min and max are increased (or decreased if the histogram is shifted to the left) by one. The above process creates an empty bin in the histogram (where the maximum bin used to be) as shown in figure

46 Chapter 3 Reversible Watermarking Based on Histogram Modification Empty bin Fig7. Shifted histogram to the right and empty bin (where the maximum bin used to be) The watermark information is later on embedded in the empty bin created by the process described above. The image is swept and whenever a pixel with luminosity value equal to max is found a watermark bit is embedded. If the watermark bit to be embedded is zero the pixel stays untouched otherwise if the watermark bit is one the pixel s luminosity is increased by one. The capacity provided by this scheme is equal to the initial input of the

47 Chapter 3 Reversible Watermarking Based on Histogram Modification maximum bin (the bin with most common luminosity). The histogram of the watermarked image is shown in figure 8. Fig8. Luminosity histogram of the watermarked image In case the capacity obtained is not sufficient for the watermark (information to be embedded) more than one maximum-minimum pairs may be selected as shown in figure

48 Chapter 3 Reversible Watermarking Based on Histogram Modification Fig9. Flowchart of the embedding procedure In order to extract the embedded information (watermark) the reverse procedure is applied. The image is swept in the same

49 Chapter 3 Reversible Watermarking Based on Histogram Modification manner and each time a pixel with luminosity value of max or max+1 is found a watermark bit is extracted. Each pixel of luminosity value max corresponds to a watermark bit of zero while each pixel of luminosity value max corresponds to a watermark bit of one. Zhicheng Ni et al s [26] watermarking scheme is fully reversible, provides high capacity (5-80 Kbits for an image of 512X512 pixels) and has low computational cost. The only disadvantage of this system is its low robustness against attacks. An improvement of [26] was proposed by Lee Sang-Kwang, Suh Young-Ho and Ho Yo-Sung [20]. In this scheme a difference image is generated as the difference of each pixel of even line from the respective pixel of the next odd line. For an image of MXN dimensions the difference image is calculated by D(i, j) = I(i, 2j + 1) - I(i, 2j), 0 i M-1, 0 j N/2-1 and has the half size of the initial image. In this case the watermark is not embedded in the luminosity histogram of the host image but in the histogram of the difference image. The difference image has the half size of the initial image. This algorithm takes advantage of the spatial correlation between

50 Chapter 3 Reversible Watermarking Based on Histogram Modification consecutive pixels. In this way the histogram of the difference image is significantly more localized with a higher maximum comparing to the maximum of the histogram of the initial image [20]. This is shown in figures 10 and 11 respectively. Fig10. Luminosity histogram for Elaine image

51 Chapter 3 Reversible Watermarking Based on Histogram Modification Fig11. Histogram of difference image for Elaine image In this scheme the capacity of the watermark is significantly increased since, as in Zhicheng Ni et al s method [26], the capacity is determined by the magnitude of the maximum of the histogram which is significantly higher, as shown in figure 10. The disadvantage of the system is that it has low robustness since it is a LSB technique

52 Chapter 3 Reversible Watermarking Based on Histogram Modification 3.2 Proposed watermarking scheme At present most watermarking schemes perform poorly against geometrical attacks [30]. The most common geometrical attacks are rotation, flipping, translation, aspect ratio changes, resizing and cropping. In many cases in order to handle geometrical attacks, watermarking schemes employ several synchronization methods. These methods usually try to identify the geometrical distortions and invert them, before the watermark detector is applied. The identification of the geometrical distortions is achieved by examining a registration pattern embedded along with the watermark in the host image [29], [7]. However the addition of the registration pattern to the data-carrying watermark reduces the fidelity of the watermarked image, as well as the scheme s capacity. Another weakness of this approach is that usually all the watermarked images carry the same registration watermark. Therefore it is easier to discern the registration watermark by collusion attempts. Once found, the registration pattern could be removed from all the watermarked images, thus restricting the invertibility of any geometric distortions. Additionally

53 Chapter 3 Reversible Watermarking Based on Histogram Modification these methods increase computational time substantially and in some cases perform poorly. There are also watermarking schemes that try to achieve robustness against geometrical attacks, using transformations, invariant to some attacks, and correlation functions like [23] and [19]. Although such schemes show good results with regard to robustness, they require high computational complexity during embedding as well as during extracting the watermark. A novel fully reversible watermarking scheme which is robust to geometrical attacks with low computational cost and is based on histogram modification was presented by Chrysochos, Fotopoulos, Skodras and Xenos [5] Features of the Proposed Scheme The proposed watermarking scheme has the following features: Reversible; the watermarked image can be fully restored to its original status

54 Chapter 3 Reversible Watermarking Based on Histogram Modification Blind; in order to detect the watermark only the watermarked image is needed. Asymmetric; a public key is used for detecting the watermark and a private key is used for restoring the watermarked image. Robust against geometrical attacks like rotation, flipping, translation, aspect ratio changes and resizing, warping, shifting, drawing, scattered tiles as well as their combinations. Multicast; a certain watermark can be embedded several times to increase robustness. Low computational cost. Applicable to very small images (down to 16 x 16). Applicable to colour images. Good watermarked image quality. The basic principle of this scheme is based on the permutation of histogram bins, which are chosen in couples, according to a specific rule. For embedding the watermark a key is

55 Chapter 3 Reversible Watermarking Based on Histogram Modification required. This is a real number that specifies the area where the watermark is to be embedded. This key is also necessary for the detection and extraction of the watermark. Therefore it is considered as public key. A second key, which is referred to, as private key, is used for the full restoration of the image Embedding Procedure The most important parameter needed, in order to embed the watermark into the host image, is the public key. This key (as mentioned above) is a real number that determines the area where the watermark is to be embedded. Its integer part (start) indicates the point of the histogram, where the embedding procedure will start choosing histogram bin couples. Its decimal part, multiplied by ten, defines the minimum distance, two histogram bins of a couple may have. We will refer to this distance as step. start = public_key div 1 step = 10*(public_key mod 1) The steps, of the embedding algorithm, in order to embed a watermark in a grayscale image are the following:

56 Chapter 3 Reversible Watermarking Based on Histogram Modification a. The histogram of the host image is computed. b. start and step are calculated with respect to the public key. c. The first couple (a, b) of the histogram bins is chosen according to start and step, as shown in figure 12. If the corresponding values of the histogram hist(a) and hist(b) are equal, we reject this couple and we continue with the next one. d. For each couple (a, b) and each bit (w) of the watermark respectively, the following rule is applied. If w equals zero then the histogram values hist(a) and hist(b) should be in ascending order. In the opposite case, i.e. when w equals one, the histogram values should be in descending order. w = 0 hist(a) < hist(b) w = 1 hist(a) > hist(b) If the values of a couple are not in the right order according to this rule, then they are swapped, in order to follow the rule. When two values of the histogram are swapped, we ensure that the value of the corresponding pixels, with luminance a and b respectively, are interchanged

57 Chapter 3 Reversible Watermarking Based on Histogram Modification hist(a) hist(b) Figure12. Host Image Histogram and corresponding couple (a, b) e. The next couple (a, b) of the histogram bins is chosen according to start and step for embedding the next bit (w) of the watermark. Steps c and d of the algorithm are repeated until all bits (w) of the watermark are embedded in the image. f. The private key, which is necessary for the restoration of the image, is generated in accordance with the procedure watermark embedding. For each couple (a, b) of the histogram bins that are chosen, a bit (pk) of the private key is generated according to this rule: if originally hist(a) and hist(b) are in

58 Chapter 3 Reversible Watermarking Based on Histogram Modification ascending order, pk equals zero. Else, if originally hist(a) and hist(b) are in descending order, pk equals zero. hist(a) < hist(b) pk = 0 hist(a) > hist(b) pk = 1 g. In case that the algorithm reaches the end of the histogram (which corresponds to intensity value of 255), it continues from the beginning of the histogram (which corresponds to intensity value of 0), provided that the couple (a, b) does not collide with a previously selected couple. In order to avoid artifacts in the watermarked image, the maximum distance between a and b is set to 9. Another case where artifacts could arise, is when a is at the end of the histogram, while b is at the beginning. Such a case is foreseen and prohibited. The flowchart of the embedding procedure is shown in figure

59 Chapter 3 Reversible Watermarking Based on Histogram Modification Host image, Watermark, Public Key Histogram generation YES End of watermark NO Watermarked Image, Private Key a, b selection Private Key generation Histogram Modification-Embedding Figure13. Embedding Algorithm

60 Chapter 3 Reversible Watermarking Based on Histogram Modification During the watermark embedding, the shape of the histogram stays almost unaltered. This is shown in figures 14 and 15 where figure 14 depicts the histogram of the original image Elaine, while figure 15 depicts the histogram of the watermarked image. Figure14. Luminosity histogram of Elaine image

61 Chapter 3 Reversible Watermarking Based on Histogram Modification Figure15. Luminosity histogram of watermarked image Only when the two histograms are more thoroughly examined, one can see in detail the changes that are induced by the embedding algorithm, as the histogram bins are interchanged. This is shown in figures 16 and 17. Figure 16 shows a close up view of the luminosity histogram of Elaine image, while figure 17 shows a close up view of the respective region of the histogram of the watermarked image

62 Chapter 3 Reversible Watermarking Based on Histogram Modification Figure16. Close-up view of luminosity histogram of Elaine image hist(a) hist(b) Figure17. Close-up view of Luminosity histogram of watermarked image

63 Chapter 3 Reversible Watermarking Based on Histogram Modification Watermark Extraction The parameters needed in order to extract the watermark from a possibly marked image, apart from the image itself, is the public key and the watermark size. The public key, as mentioned above is a real number that determines the histogram area where the watermark is embedded in. The scheme is blind and thus there is no need for the original image. The steps, of the extracting algorithm, in order to extract the watermark out of a grayscale image are the following: a. The histogram of the watermarked image is computed. b. start and step are calculated with respect to the public key. c. The first couple (a,b) of the histogram bins is chosen according to start and step. If the corresponding values of the histogram hist(a) and hist(b) are equal, this couple is rejected and the process continues with the next one. d. Each couple (a,b) corresponds to a single bit (w) of the watermark. The following rule is applied. If the histogram values (hist(a), hist(b)) are in ascending order w equals zero. In the opposite case, where histogram values are in descending order w equals one

64 Chapter 3 Reversible Watermarking Based on Histogram Modification hist (a) < hist (b) w = 0 hist(a) > hist (b) w = 1 e. The next couple (a, b) of the histogram bins is chosen according to start and step for extracting the next bit (w) of the watermark. Steps c and d of the algorithm are repeated until all the bits (w) of the watermark are extracted. It is crucial for the extracting process, to choose the same couples (a, b), that were chosen during the embedding process. Therefore the choice of the couples follows exactly the same rules with the embedding algorithm. The process of extracting the watermark can be successful only if the couples that are chosen for extraction are exactly the same with those, chosen for embedding. The flowchart of the extracting algorithm is shown in figure

65 Chapter 3 Reversible Watermarking Based on Histogram Modification Watermarked image, Public key Histogram generation YES End of watermark NO Watermark a, b selection Watermark Extraction Figure18. Extracting Algorithm

66 Chapter 3 Reversible Watermarking Based on Histogram Modification Restoration of Watermarked Image In order to restore the watermarked image to its original form, both the private and the public keys are needed. The private key is the key that has been created during the embedding process. The steps of the restoring algorithm, in order to retrieve the original image, are the following: a. The histogram of the watermarked image is computed b. start and step are calculated with respect to the public key. c. The first couple (a,b) of the histogram bins is chosen according to start and step. If the corresponding values of the histogram hist(a) and hist(b) are equal, this couple is rejected and the algorithm proceeds to the next one. d. For each couple (a,b) and each bit (pk) of the private key respectively, the following rule is applied. If pk equals zero then the histogram values (hist(a), hist(b)) should be in ascending order. In the opposite case, where pk equals one, the histogram values should be in descending order. pk = 0 hist (a) < hist (b) pk = 1 hist (a) > hist (b)

67 Chapter 3 Reversible Watermarking Based on Histogram Modification If the values of a couple are not in the right order according to this rule, then they are swapped, in order to follow the rule. When two values of the histogram are swapped, the intensity values of the corresponding pixels are interchanged. e. The next couple (a,b) of the histogram bins is chosen according to start and step. Steps c and d of the algorithm are repeated until all the bits (pk) of the private key are examined with respect to the histogram. It is crucial for the restoring process to choose the same couples (a, b), that were chosen during the embedding process. Therefore the choice of the couples follows the exact same rules with the embedding algorithm. 3.3 Experimental results The watermarking algorithm that was described in section 3.2 could be also applied for the case of colour pictures. The only difference is that instead of gray scale intensity values, it should be

68 Chapter 3 Reversible Watermarking Based on Histogram Modification applied to the color components, RGB, or YCbCr respectively. In this way a watermark could be embedded three times, thus achieving improved robustness. The robustness in geometrical attacks is increased, as a change in a pixel affects different bins of each component s histogram, thus it is less probable that all three watermarks are affected in the same way. Furthermore, robustness may increase more by embedding a symmetrical watermark, in a way that allows integrity check, as well. So, if the extracted watermark follows a given symmetry, it is ensured that the watermark is intact. On the other hand, if increased capacity is of prime concern, a three times larger watermark can be embedded partially in each color component. The maximum capacity of this scheme is rather low, namely 128 bits, but it may rise to 384 bits, if we use three color components. It has the advantage, though, that it can be applied to very small images (down to 16 x 16), with reduced capacity. The scheme may provide up to 2304 different public keys (this results as the combination of 256 different values for start with 9 different values for step) for gray scale images and different public keys (2304^3) for colour images