An image adaptive, wavelet-based watermarking of digital ... - CiteSeerX

ARTICLE IN PRESS

Journal of Computational and Applied Mathematics

(

)

– www.elsevier.com/locate/cam

An image adaptive, wavelet-based watermarking of digital images Santa Agrestea , Guido Andalorob , Daniela Prestipinob , Luigia Pucciob,∗ a Department of Information Science, University of Milano, Italy b Department of Mathematics, University of Messina, Italy

Received 19 September 2005; received in revised form 30 May 2006

Abstract In digital management, multimedia content and data can easily be used in an illegal way—being copied, modified and distributed again. Copyright protection, intellectual and material rights protection for authors, owners, buyers, distributors and the authenticity of content are crucial factors in solving an urgent and real problem. In such scenario digital watermark techniques are emerging as a valid solution. In this paper, we describe an algorithm—called WM2.0—for an invisible watermark: private, strong, waveletbased and developed for digital images protection and authenticity. Using discrete wavelet transform (DWT) is motivated by good time-frequency features and well-matching with human visual system directives. These two combined elements are important in building an invisible and robust watermark. WM2.0 works on a dual scheme: watermark embedding and watermark detection. The watermark is embedded into high frequency DWT components of a specific sub-image and it is calculated in correlation with the image features and statistic properties. Watermark detection applies a re-synchronization between the original and watermarked image. The correlation between the watermarked DWT coefficients and the watermark signal is calculated according to the Neyman–Pearson statistic criterion. Experimentation on a large set of different images has shown to be resistant against geometric, filtering and StirMark attacks with a low rate of false alarm. Published by Elsevier B.V. MSC: 65T60; 65Y99; 68U10; 94A08; 94A62 Keywords: Copyright protection; Digital watermark; Image processing; Wavelet transform

1. Introduction In digital management, multimedia content and data can easily be used in an illegal way, this is easier than ever before. The large availability of multimedia data in digital format and its distribution on Internet has in fact not changed, with respect to a traditional context, the relevant legal rights but rather the capacity and the simplicity of violation of the same rights [7,4,15]. Audio, video, images in digital format are subject to risk of tampering and violation of copyright: they can easily be copied, manipulated and distributed again. Copyright protection, data authenticity, control on unauthorized copying and distribution of digital objects have become a crucial issue solving an urgent and real problem for owner, sellers, vendors, buyers and customers. ∗ Corresponding author.

E-mail addresses: [email protected] (S. Agreste), [email protected] (G. Andaloro), [email protected] (D. Prestipino), [email protected] (L. Puccio). 0377-0427/$ - see front matter Published by Elsevier B.V. doi:10.1016/j.cam.2006.10.087 Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087

ARTICLE IN PRESS 2

S. Agreste et al. / Journal of Computational and Applied Mathematics

(

)

–

Fig. 1. (a) Embedding scheme: original watermark signal WS is embedded in the original image OI, the output is the watermarked image WI. (b) Detection scheme: watermark signal WS∗ is extracted from the watermarked image WI∗ (likely different from WI) and then compared with the original watermark signal WS by means a correlation function F and a fixed threshold .

The possibilities offered intrinsically by digital media to control copyright and rightful utilization result dramatically minimized, so as to involve digital right management (DRM) systems. In a DRM framework various complimentary tools are used solving different problems, these involve both public key infrastructure and information hiding techniques. Copyright protection and data authenticity as well as rightful utilization of it and its traceability represent, today and in the next future, issues of substantial interest to which the international scientific community will have to give an answer. Therefore it is strategic to individuate and to develop methods and numerical algorithms, which are stable and with low computational cost, that will allow to answer these questions. To this aim the tools offered by the literature are the following: watermarking, fingerprinting and cryptographic schemes. In this paper we put the focus on watermarking techniques and describe the realized wavelet-based algorithm for watermarking of still images. The rest of the paper is organized as follows: in Section 1 we present an overview about tools and motivation, in Section 2 we describe in detail the proposed algorithm, in Section 3 we present some experimental results, Section 4 concludes this work. 1.1. Watermark properties Digital watermark is a visible or invisible identification code that is permanently embedded in the host media. Invisible watermark is an imperceptible copyright information which is hidden directly in media content in such a manner that it cannot be removed by the user but it can be extracted or read by the appropriate party and under specific conditions. Watermark of the media aims at discouraging unauthorized copying. Alterations applied to the content by watermark embedding must be imperceptible, and quality of the content must not decrease after watermarking, which must be robust to media manipulations, tampering and attacks. In detail, the watermark is an interference signal inserted into a media data signal. Watermark of digital images must be: invisible, robust to attacks and tamper-resistant to legitimate signal processing operations [5]; it must have a low rate of false alarm and a high level of secrecy to generate the watermark information. A typical formulation of watermarking process counts two schemes: embedding and detection. In the former the goal is encoding the watermark signal into an image, whereas in the latter it checks the boolean presence of the watermark signal into watermarked image computing the correlation between the watermarked components of the image (which can likely be changed) and the original watermarking signal, by means of a correlation function and a fixed threshold (Fig. 1). Watermark algorithm can be classified into private (not blind) and public (blind) depending on the requirement of the original image during the detection process. Private schemes require the original image during the detection process, but are more robust than public schemes (Fig. 1). Another classification criterion distinguishes watermark schemes into spatial domain techniques and transform domain techniques in relation to approach followed to process the original image. In the former the watermark is encoded by directly modifying pixels, whereas in the latter the watermark is encoded by altering some frequency bins obtained by transforming the image in the frequency domain [5]. Spatial domain techniques are less complex, no transform is involved but they are less robust to tampering and attacks than transform domain techniques which place the watermark signal in the most perceptually significant components of a transform domain (Fourier, wavelet, cosine) [3,13,8]. Stego image data can be altered by various signal processing or geometric operations. Attacks to watermarked digital images can be classified as presented in [17]: removal attacks, geometrical attacks, cryptographic attacks and protocol attacks. Removal attacks remove the watermark from the watermarked image, they include: denoising, lossy compression, quantization, remodulation, collusion and averaging attacks. Geometrical attacks do not remove the embedded watermark but distort it through spatial or temporal alterations of the watermarked image, such that the detection process is not synchronized with the watermark signal embedded. Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087

ARTICLE IN PRESS S. Agreste et al. / Journal of Computational and Applied Mathematics

(

)

–

3

Fig. 2. (a) Lena: original image. (b) Lena: two-level decomposition of DWT by Daubechies 2. (c) Lena: two-level decomposition of DWT by Daubechies 2, coefficient distribution. (d) DWT with three-level of decomposition.

Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087

ARTICLE IN PRESS 4

S. Agreste et al. / Journal of Computational and Applied Mathematics

(

)

–

Cryptographic attacks are the brute force attacks which aim to find the secret information used to generate the watermark signal. Finally the protocol attacks aim at attacking the concept of the watermarking application [17]. Attacks are the cause of detection errors: false positive error (FPE) and false negative error (FNE). FPE detects the watermark when it is not embedded, FNE does not detect the watermark when it is embedded. 1.2. Watermarking in the wavelet domain The wavelet functions well analyze image features such as edges and borders thanks to a good space-frequency localization. The main property of wavelet functions is to process data at different scales or resolutions, highlighting both large and small features. Wavelet functions can process signals containing many discontinuity or sharp changes. They are used in several fields: image compression, signal denoising, image smoothing and texture analysis. Wavelet functions make the watermarking scheme more robust in comparison with spatial methods because of irregular distribution of the inverse transform value, which makes the watermark signal distributed over the image [14]. The main advantages of inserting watermarks in the wavelet transform domain can be found in [2,9,10] and in references therein. Space-frequency localization: image features are well analyzed by wavelet domain space-frequency localization, this property makes the watermark most robust to the geometric attacks. Multi-resolution representation: it allows to detect hierarchical process, this property is very important for watermark detection. Superior HVS modeling: watermarking benefits from a human visual system (HVS) model, in fact matching HVS directives watermark interference can be masked to properties and sensibility of the human eye. Watermark signal can be embedded into the three largest detailed sub-bands, with the selection of watermark strength based on the local sensitivity of the image to noise. Linear complexity: in wavelet domain watermarking requires a lower computational cost O(n) than Fourier and cosine O(n log(n)). Adaptivity: wavelet and related filters can be chosen with a dynamic mode, reflecting the properties of the image. Finally it is important to observe compatibility to the upcoming image coding standard JPEG 2000 (Fig. 2). 2. Watermarking algorithm The realized algorithm, called WM2.0, is a watermarking not blind algorithm, which embeds watermark signals into high-frequency sub-bands discrete wavelet transform (DWT) coefficients, according to the HVS directives [6]. It makes a pre-processing of the image depicting it into component value of color model hue, saturation, value (HSV) and resizing the value matrix in accordance with the parameters and mathematical base conditions of DWT. Wavelet function and DWT level decomposition are fixed, respectively, depending on image features and image resize. In the embedding process, watermark signal and DWT coefficients to be watermarked are chosen depending on the statistic function values of the image. In the detection process, original image and watermarked image (likely different from the output image of the embedding process because of JPEG compression or any attacks) are synchronized comparing statistic function values of a geometric interval of both images; the correlation between the watermarked DWT coefficients and the watermark signal is calculated according to the Neyman–Pearson statistic criterion which determines a detection threshold minimizing the probability of missing detection to a given probability of false alarm. WM2.0 is an evolution of a previous algorithm version, WM1.0, described in [1]. In WM1.0 watermark signal and detection threshold were constant values chosen by means of experimental considerations, thus they were not depending on statistic image features. The experimentation has been accomplished on images, in high and low resolutions, building a real and commercial database. This algorithm has been implemented in Matlab 6.x using the wavelet and statistic toolbox. 2.1. Pre-processing of the image It is widely accepted today that robust image watermarking techniques should largely exploit the characteristics of the HVS, for more effectively hiding a robust watermark [5,16]. HVS considerations indicate that the eye is less sensitive to noise in those areas of the image where brightness is high or low. For this reason, in this step, first we compute the value plane from HSV model of the original image I . Three fundamental considerations indicate us to apply the wavelet transform not on the whole value matrix of the original image, but on its sub-matrices: the first reason is that, in this way, the watermark can well cover the whole image; the second reason is that DWT requires that the associated sub-image matrix must have order power of 2; and the third reason is that the host images can have different dimensions when they belong to real multimedia galleries. Then, in this step, we split value matrix of the original image Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087

ARTICLE IN PRESS S. Agreste et al. / Journal of Computational and Applied Mathematics

(

)

–

5

Fig. 3. (a) The value matrix of the host image, which to be watermarked, is dipped in original image. (b) The original image is dipped in value matrix associated to the block. (c) More blocks cover the original image.

I , into non-overlapping squared blocks. It is important to define a criterion to compute the blocks to be watermarked. In WM2.0 we apply the following rule, let r be the number of rows and c be the number of columns of I : • if c/r < 2, then each block is a matrix whose order power of 2 is nearer to the longest dimension of the original image (typically, we have one block, Fig. 3(a) or (b)). If the order of the matrix associated to the block is greater than the longest dimension of the original image, we apply the pixel values of the first and last rows, and of the first and last columns; • if c/r 2, then each block is a matrix whose order power of 2 is nearer to the lowest dimension of the original image (typically, we have more blocks, Fig. 3(c)). In this way, on each block we apply the embedded/detection scheme described in the next sub-sections. In experimentation cases typically computed have been of one block with an associated matrix with order 256, or more blocks with an associated matrices with order 128. 2.2. Watermark embedding Let C be the matrix associated to block to be watermarked. C has order n power of 2. In the watermark embedding step DWT decomposition is applied to C to obtain the four sub-matrices CkLL , CkH L , CkLH , CkH H of order nk = n/2k , which have the DWT coefficients of the kth decomposition level as elements (Fig. 2(d)). Only the entries of high frequencies detail matrices Ck , where  ∈ {H L, LH , H H } are modified by watermark. The number of DWT wavelet decomposition depends on the order of sub-image matrix associated to blocks. Typically we apply three DWT decomposition levels on a block with associated matrix of order 256, and two DWT decomposition levels on each block with associated matrix of order 128. Watermark embedding is calculated with the following formulas: C˜ k (i, j ) = Ck (i, j ) +  ∗ (i, j ),

(1)

where i, j = 1, 2, . . . , nk and (i, j ) is the generic element of a matrix of order equal to the order of Ck . (i, j ) ∈ {−1, 0, 1} and its value is computed depending on the belonging of the corresponding DWT coefficient to an interval. The interval depends on false alarm probability Pf and , which is equal to standard deviation (STD) of the DWT coefficients. If we fix a priori the variance 2 , defined in (3), and the probability Pf in (4) we have ⎞ ⎛  (2) Ck ⎠ ,  = STD ⎝ 

2 = 1/(3N)2

nk nk  

 i=1 j =1

Ck (i, j )2 ,

  T = erfc 2 ∗ Pf ∗ 22 ,

(3)

(4)

Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087

ARTICLE IN PRESS 6

S. Agreste et al. / Journal of Computational and Applied Mathematics

(

)

–

Fig. 4. (a) Lena: value (HSV) plane of the original image. (b) Lena: difference between the original (value) matrix and the watermarked matrix. (c) Lena: watermark matrix  in related level decomposition. (d) Lena: DWT coefficients before and after watermark embedded.

where erfc denote the error function. Then the value of parameter (i, j ) is • (i, j ) = 0, if − < Ck (i, j ) < ; • (i, j ) = 1, if Ck (i, j ) > T ; • (i, j ) = −1, otherwise. The inverse discrete wavelet transform (IDWT) and inverse selection scheme are computed on C, so as to have HSV of the watermarked image, and storage in JPEG format for distribution completes the embedding scheme. JPEG Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087

ARTICLE IN PRESS S. Agreste et al. / Journal of Computational and Applied Mathematics

(

)

–

7

Fig. 5. (a) Lena: watermarked image without attacks, plot of  and T . (b) Lena: watermarked image with StirMark attack, plot of  and T . (c) Lena: watermarked image with geometric attack (cutout), plot of  and T .

compression has a quality factor which minimizes the peak signal to noise ratio (PSNR). We denote with I˜ the watermarked image. 2.3. Watermark detection This step checks if original and watermarked image sizes are equal: if they are different a synchronization process computes a central block of the original image with respect to recognize some blocks of the watermarked image by means of the mean square error (MSE) function. On watermarked image I˜ (which can likely be attacked) and original image I the same steps from pre-processing to DWT decomposition are computed. Watermark is detected computing the correlation between the watermarked coefficients and watermark signal, in comparison to the threshold T : ⎡ ⎤ nk nk    ⎣  = 1/3N C˜ k (i, j ) − Ck (i, j )⎦ , (5) i=1 j =1





 Pf 1/2 erfc T / 22 ,

(6)

Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087

ARTICLE IN PRESS 8

S. Agreste et al. / Journal of Computational and Applied Mathematics

where 2 is computed by (3), Pf is fixed to 10−8 and T = erfc(2 ∗ Pf ∗ detected, otherwise, watermark signal is not detected.

(

)

–

√ 22 ), then if  > T watermark signal is

3. Attacks and experimental results The experimentation, experimental tests on false alarms and attacks have been accomplished on 1000 images, in high and low resolutions, building a real and commercial database. Experimentation results have shown that WM2.0 is robust against attacks of geometric operation, filters and StirMark [11,12], with a ratio of more than 88%. It has a low probability of false positive alarm, and processes the image without false negative (Figs. 4, 5). In particular, Fig. 5 shows the values of  and T for different choices of Pf from 10−15 to 10−5 . WM2.0 processes a still image, with a format of 350 × 350 pixels and with a file dimension of 35 Kb, in 2.97 s for embedding and 2.78 s for detection. Elaboration time depends on blocks subdivision of the image and thus from DWT levels decomposition. 4. Conclusion The watermark embedded has a high level of robustness against geometric and image processing attacks and a low rate of false alarm. Using DWT on HSV color model, image statistic features and image pre-processing step with blocks subdivision of the image are key steps in this algorithm for robustness, invisibility and a signal watermark distribution over image. Acknowledgments Experimentation was supported by two applied research projects: Ecumene and Beni Ecclesiastici in Web, according to owner of images which is Italian Episcopal Conference, and relationship with publisher of the images—I.D.S. Informatica—an ICT company located in Messina (Italy). Ecumene is a project funded within project Parnaso of MIUR, the Italian Ministry of Education, University and Research. Beni Ecclesiastici in Web is a multimedia project of CEI, the Italian Episcopal Conference. References [1] S. Agreste, N. Castorina, S. Giovinazzo, D. Prestipino, L. Puccio, Tutela del diritto di proprieta’ delle immagini digitali: Implementazione di un algoritmo di Watermark mediante funzioni Wavelet, Atti dell’Accademia Peloritana dei Peric., Classe di Scienze Fis. Mat. e Nat., vol. LXXXI–LXXXII, 2005, pp. 1–15, ISSN 1825-1242 C1A0401009 ID Code 261. [2] M. Barni, F. Bartolini, A. Piva, Improved wavelet-based watermarking through pixel-wise masking, Image Process. IEEE Trans. 10 (5) (2001) 783–791. [3] F.M. Boland, J.J.K. Ruanaidh, W.J. Dowling, Watermarking digital images for copyright protection, IEEE Proc. Vision, Signal and Image Process. 143 (4) (1996) 250–256. [4] P. Bonne, Copyright protection and copy control when distributing and publishing digita information, GSEC Practical Version 1.4b, Option 1, SANS Institute, 2003. [5] I. Cox, M.L. Miliier, A review of watermarking and the importance of perceptual modeling, in: Proceedings of SPIE, vol. 3016, 1997, p. 929. [6] Z. Dawei, C. Guanrong, L. Wenbo, A chaos-based robust wavelet domain watermarking algorithm, Chaos, Solitons and Fractals 22 (2004) 47–54. [7] C. De Paoli, Digital right management: Cenni Storici e Principali Standard, ICT Security, Marzo 2004, pp. 72–82. [8] E. Koch, J. Zhao, Towards robust and hidden image copyright labeling, in: Proceedings of IEEE Workshop of Nonlinear Signal and Image Processing, Halkidiki, Greece, 1995. [9] A. Lumini, D. Maio, A wavelet-based image watermarking scheme, in: Information Technology: Coding and Computing, 2000, Proceedings, International Conference on 27–29 March 2000, pp. 122–127. [10] P. Meerwald, A. Uhl, A survey of wavelet-domain watermarking algorithms, in: Proceedings of SPIE, Electronic Imaging, Security and Watermarking of Multimedia Contents III, vol. 4314, CA, USA, January 2001, pp. 505–516. [11] F.A.P. Petitcolas, Watermarking schemes evaluation, IEEE Signal Process. 17 (5) (2000) 58–64. [12] F.A.P. Petitcolas, R.J. Anderson, M.G. Kuhn, Attacks on copyright marking systems, in: D. Aucsmith (Ed.), Information Hiding, Second International Workshop, IH98, Portland, Oregon, USA, Proceedings, Lecture Notes in Computer Science, vol. 1525, Springer, Berlin, April 15–17, 1998, pp. 219–239, ISBN 3-540-65386-4. [13] I. Pitas, A method for signature casting on digital images, in: Proceedings of the International Conference on Image Processing, vol. 3, 1996, pp. 215–218. Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087

ARTICLE IN PRESS S. Agreste et al. / Journal of Computational and Applied Mathematics

(

)

–

9

[14] A. Reddy Adhipathi, B.N. Chatterji, A new wavelet based logo-watermarking scheme, Pattern Recognition Lett. 26 (2005) 1019–1027. [15] M. Stefik, Letting loose the light: igniting commerce in electronic publication, Internet Dreams: Archetypes, Myths, and Metaphors, MIT Press, Cambridge, MA, 1996. [16] A.H. Tewfik, M. Swanson, Data hiding for multimedia personalization, interaction, and protection, IEEE Signal Process. Mag. 14 (1997) 41–44. [17] S. Voloshynovskiy, S. Pereira, V. Iquise, T. Pun, Attack modelling: towards a second generation watermarking benchmark, Signal Process. 81 (2001) 1177–1214.

Please cite this article as: S. Agreste, et al., An image adaptive, wavelet-based watermarking of digital images, J. Comput. Appl. Math. (2006), doi: 10.1016/j.cam.2006.10.087