This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. ??, NO. ?, APR 2011 1 Blind Image Watermarking Using a Sample Projection Approach Mohammad Ali Akhaee*, Member, IEEE, Sayed Mohammad Ebrahim Sahraeian, Student Member, IEEE, and Craig Jin, Member, IEEE Abstract—This paper presents a robust image watermarking scheme based on a sample projection approach. While we consider the human visual system in our watermarking algorithm, we use the low frequency components of image blocks for data hiding to obtain high robustness against attacks. We use four samples of the approximation coefficients of the image blocks to construct a line segment in the 2-D space. The slope of this line segment, which is invariant to the gain factor, is employed for watermarking purpose. We embed the watermarking code by projecting the line segment on some specific lines according to message bits. To design a maximum likelihood decoder, we compute the distribution of the slope of the embedding line segment for Gaussian samples. The performance of the proposed technique is analytically investigated and verified via several simulations. Experimental results confirm the validity of our model and its high robustness against common attacks in comparison with similar watermarking techniques that are invariant to the gain attack. Index Terms—Image watermarking, Maximum Likelihood detector, sample projection, gain attack. I. I NTRODUCTION Digital watermarking embeds information within a digital work as a part of the media. Watermarking techniques falls into three categories of robust, semi fragile, and fragile methods according to their specific applications [1]–[3]. Robust watermarking mainly serves for identification purposes while the fragile and semi fragile watermarking are usually employed in authentication applications. Since a good watermarking scheme should always be able to deal with some kinds of attacks, studies in the watermarking research area mostly target robust watermarking problems. Several robust watermarking techniques have been proposed so far. Cox et al. [4] have proposed an additive watermarking approach based on spread spectrum concept which remains highly robust against noise and cropping attacks. Several other studies have improved this approach further [5]–[10]. Based on this observation that boosting the watermarking power increases the barrier against attacks, most of the effective watremarking schemes try to match the characteristics of the watermark to those of the image asset. Multiplicative Copyright (c) 2010 IEEE. Personal use of this material is permitted. However, permission to use this material for any other purposes must be obtained from the IEEE by sending a request to [email protected]. M. A. Akhaee and Craig Jin are with the Computing and Audio Research Laboratory (CARLAB), Department of Electrical and Information Engineering, University of Sydney, NSW, Australia, 2006 (e-mail: akhaee, [email protected]). S.M.E. Sahraeian is with the Genomic Signal Processing (GSP) Lab, Department of Electrical and Computer Engineering, Texas A&M University, College Station, TX 77843, USA (e-mail: [email protected]). watermarking, as an example, has been introduced in [11] and has been widely studied later on using local optimum decoders in multiresolutional transform domains such as wavelet and contourlet domains [12]–[16]. Besides, a universal optimal detector for scaling based watermarking schemes is presented in [17]. These schemes are highly robust against noise and compression attacks. To satisfy robustness against geometric attacks and reduce the watermark synchronization problem, Tang and Hang [18] proposed a watermarking scheme which employs a feature extraction and image normalization approach. Based on log polar mapping (LPM) and phase correlation, Zheng et al. [19] proposed an image watermarking technique which embeds the watermark into the LPMs of the image Fourier magnitude spectrum. This scheme is invariant to rotation and remains comparatively robust to scaling attack. One of the common but effective attacks in watermarking systems is volumetric distortions ( i.e., any kind of amplitude scaling or gamma compensation). For image signals, it may happen due to the scanning process where light is not distributed uniformly over the paper. This simple attack is the main drawback of most recent studies like [12], [14]–[16] and even the quantization index modulation (QIM) algorithm [20] which has attained great popularity due to its lossless performance with lattice-based codebooks. Three types of solutions can tackle this problem, especially for the QIM method: 1) adopting auxiliary pilots through the watermarked signal known at both the encoder and decoder [21]; 2) using spherical codewords [22] with correlator decoders [23], or using angle QIM (AQIM) [24], [25]; iii) introducing a domain in which the embedding process is invariant to the gain attack [26]. Some improvement/generalization on this scheme which is referred to as rational dither modulation (RDM) are proposed in [27]– [29]. The first solution against the gain attack decreases the security of the algorithm, since the malicious attacker can change either the watermark or the pilot signals. Besides, pilots are deterministic objects in the main signal and can be easily detected. Although the second and third approach keep the security of the QIM algorithm, they cause high computational cost. Besides, the low robustness of AQIM against additive white Gaussian noise (AWGN) and the high peak to average power ratio (PAPR) of RDM algorithm which happens due to its momentarily large quantization step size are the main drawbacks that should be addressed. Thus, no approach has been presented so far that both proposes an optimal decoder and remains invariant to the gain attack. In this paper, we propose a novel gain invariant watermark- Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. 2 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. ??, NO. ?, APR 2011 ing scheme based on a sample projection scheme. Embedding the watermark bits into the approximation coefficients of the image blocks makes the algorithm highly robust against noise and compression attacks. Any possible selection of four approximation coefficients, that may be selected using a secret key, constructs a line segment whose slope is considered for data hiding. We embed the watermark bits by projecting the line segment on some specific coding lines while preserving its center of mass. In this way, the slope of the line segment carries the watermark information while the distortion imposed to its constructive samples is minimal. Since our embedding process is linear, it can be denoted by multiplication of specific embedding matrices. To implement the maximum likelihood (ML) detector for data extraction, we should calculate the distribution of the slope of the line segment. To this aim, we consider the fact that the approximation coefficients of most of the image blocks can be well-modeled by Gaussian distribution [15]. The performance of our watermarking method is analytically investigated and verified via simulation on artificial signal. Several experiments on sample images confirm the effectiveness of the proposed scheme in resisting against common noise attacks. The rest of the paper is organized as follows. In Section II, we describe the model of the system. The watermark embedding and decoding process are introduced in Section III. Section IV analyzes and evaluates the performance of the proposed scheme. Experimental results are demonstrated in Section V, and Section VI concludes the paper. II. S YSTEM M ODELING In this section, we first introduce the model considered for our watermarking algorithm. To this aim, we calculate the distribution of the watermarking variable. We assume to have four samples of an independently and identically distributed (i.i.d) Gaussian random variable as the host signal. We show this signal as u = [u1 , u2 , u3 , u4 ] with the Gaussian distribution of N (0, σu2 ). These four samples form two points p = [u1 , u2 ] and q = [u3 , u4 ] in the 2-D space. We employ 2 c = uu43 −u −u1 , the slope of the line going through these two points as our watermarking variable. We can see that the numerator and the denominator of c are distributed as N (0, 2σu2 ). For independent Gaussian variables a ∼ N (0, σa2 ) and b ∼ N (0, σb2 ), their ratio c = ab which is the ratio of two zero-mean independent Gaussian variables is Cauchy as: σ fC (c) = a 1 σb , π c2 + ( σσab )2 (1) As shown in Appendix A, for the case that the two Gaussian variables a and b are correlated with the correlation coefficient r= E[(a − µa )(b − µb )] , σa σb the probability density function (PDF) of the variable c is given as: √ σa σb 1 − r 2 1 , (2) fC (c) = a 2 2 2 π σb2 (c − rσ σb ) + σa (1 − r ) and its cumulative distribution function (CDF), FC (c), can be computed as: FC (c) = 1 1 σb c − rσa + tan−1 √ . 2 π σa 1 − r 2 (3) As we will see in the next sections, we use the probabilistic characteristics of the variable c, the slope of the mentioned line segment, to design an optimal decoder and later analyze its performance. III. P ROPOSED M ETHOD In this section, we introduce our blind watermarking scheme. As discussed in the previous section, we assume the host signal as a four-sample i.i.d. Gaussian random signal. In practical applications, these four samples can come from approximation coefficients of the image blocks which satisfy our i.i.d. Gaussian assumption according to KolmogorovSmirnov test results. A. Watermark embedding Let us represent the four samples of the host signal as u = [u1 , u2 , u3 , u4 ]. We model these four samples as two points p = [u1 , u2 ] and q = [u3 , u4 ] in the 2-D space. Fig. 1(a) illustrates these two points as well as the line segment connecting them. We denote the slope of this line segment as 3 u2 +u4 θ. The center of this (p, q) line is located at [ u1 +u , 2 ]. 2 By translating the center of this line to the origin, we reach to the points pc and qc , where ( ′ ) ( u1 −u3 ) ( ′ ) ( u3 −u1 ) u1 u3 2 2 pc = = u2 −u4 , qc = = u4 −u . (4) 2 u′2 u′4 2 2 Now, to embed the M -ary watermark code, we project this line segment to one of the M coding lines (L1 to LM if θ > 0 or their counterparts L′1 to L′M if θ < 0), shown in Fig. 1 (b), depending on the watermark code. We use projection and keep the center of the line segment to impose less distortion and thereby cause more invisibility of the watermarked signal. Fig. 1 (a) illustrates the projection steps in details. We call the resulted points in the mapped line as p⊥ and q⊥ . The slope of the ith coding line is αi (for Li ) if θ, the slope of the primary line connecting (p, q), is positive and −αi (for L′i ) otherwise, where αi is given by π M − 2i + 1 − β). (5) 4 2 Here, β is the angle between two consecutive coding lines. The position of p⊥ for the case that the slope of the mapped line is k, can be computed using the intersection of two following lines: { y = kx (6) y − u′2 = − k1 (x − u′1 ). αi = tan( With a similar scheme, we can find q⊥ . The obtained points are as follows: ( u′ +ku′ ) ( u′ +ku′ ) 1 p⊥ = 2 k2 +1 ku′1 +k2 u′2 k2 +1 3 , q⊥ = 4 k2 +1 ku′3 +k2 u′4 k2 +1 . (7) Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. IMAGE WATERMARKING 3 u′′1 u1 u′′2 u2 ′′ = T(k) , u3 u3 ′′ u4 u4 as: (9) using the transfer matrix T(k), given as: 2 k +2 k 1 T(k) = 2 2k + 2 k 2 −k k k2 −k 2k 2 + 1 −k 1 . (10) −k, k2 + 2 k 1 k 2k 2 + 1 Therefore, the watermarking embedding process can be figured as implementation of (9) where k is defined based on the watermark bit and θ, the slope of the primary (p, q) line, as: { αi when θ ≥ 0 k= (11) −αi when θ < 0. (a) In Fig. 1(b), we can see that for each coding word we have two lines, a solid one (Li ) for positive θ and a dashed one (L′i ) for negative θ. In this way, we obtain the watermarked signal u′′ = ′′ [u1 , u′′2 , u′′3 , u′′4 ]. B. Watermark decoding (b) Fig. 1. (a) Steps of the proposed watermarking embedding scheme: translation to the origin, projection, and translation back. (b) Coding space. Solid lines: coding lines with positive slopes (L1 · · · LM ); dashed lines: the counterpart coding lines with negative slopes (L′1 · · · L′M ); dotted lines: Decision boundaries. As the final step, we just need to translate back the mapped line to its original center to obtain points pw = [u′′1 , u′′2 ] and qw = [u′′3 , u′′4 ] as: ) ( ′′ ) ( u′1 +ku′2 u1 +u3 u1 k2 +1 + 2 pw = = ku′ +k2 u′ 1 2 4 u′′2 + u2 +u k2 +1 2 ) ( ( ′′ ) u′3 +ku′4 u1 +u3 u3 k2 +1 + 2 . qw = = ku′ +k2 u′ 3 4 4 u′′4 + u2 +u k2 +1 2 To extract the hidden bits, an optimum decoder is implemented using M-Hypothesis test as follows. We denote the received signal as y = u′′ + n, where y = [y1 , y2 , y3 , y4 ] represents the watermarked signal contaminated with zero mean AWGN n ∼ N (0, σn2 ). For the i.i.d. Gaussian host signal u, the watermarked signal u′′ is also Gaussian as it is obtained by a linear transformation through the matrix T(k). Thus, the received samples y are Gaussian with the variance of σy2 = σu′′2 + σn2 . Now, we can use the model given in Section II. We depict the four samples in y as two points pr = [y1 , y2 ] and qr = [y3 , y4 ] in the 2-D space and calculate the slope of the line segment connecting them as: y4 − y2 u′′ − u′′2 + n4 − n2 c= = 4′′ . (12) y3 − y1 u3 − u′′1 + n3 − n1 As we fixed the slope of the line connecting pw = [u′′1 , u′′2 ] and qw = [u′′3 , u′′4 ] to k, defined in (11), we have: k= u′′4 − u′′2 ⇒ u′′4 − u′′2 = k(u′′3 − u′′1 ). u′′3 − u′′1 (13) Therefore, by defining v = u′′3 − u′′1 , we can rewrite (12) as: c= kv + n4 − n2 . v + n3 − n1 (14) Using (9) and (10), we have 1 (−u1 − ku2 + u3 + ku4 ). +1 Thus, v is a Gaussian random variable with mean µv = 0 2 2 and variance σv2 = 1+k 2 σu . Consequently, if we show the numerator and the denominator of c in (14) respectively as a and b, we will have a ∼ N (0, k 2 σv2 +2σn2 ) and b ∼ N (0, σv2 + v = u′′3 − u′′1 = (8) By inserting (4) in (8) we can summarize the whole procedure k2 Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. 4 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. ??, NO. ?, APR 2011 2σn2 ). The correlation coefficient between a and b can also be computed as: r=√ kσv2 (k 2 σv2 + 2σn2 )(σv2 + 2σn2 ) . can be written as: P + (e|i) =P (tan (−ϕi + β β ) < c|i) + P (c < tan (−ϕi − )|i) 2 2 ( ( β ) β ) =FC|i tan (ϕi − ) − FC|i tan (−ϕi + ) 2 2 ( ( β ) β ) +1 − FC|i tan (ϕi + ) + FC|i tan (−ϕi − ) , 2 2 (21) (15) +P (tan (ϕi + Now we have c = ab , where a and b are two correlated zero mean Gaussian random variables. Thus, the distribution function of c can be given as in (2). Having the distribution of the slope function, we can simply use the Maximum Likelihood detector as: i∗ = arg max fC (c|i), (16) i∈1,··· ,M where fC (c|i) represents the distribution function for the ith coding line. Here, without loss of generality, we suppose that the received c is positive. The same argument also holds for the negative case. Thus, the conditional distribution function is √ σ σ 1 − r|i2 a b |i |i 1 fC (c|i) = , (17) π σ 2 (c − r|i σa|i )2 + σa2 (1 − r2 ) b|i σb|i |i where, c is the slope of the detected line, and FC|i (.) is its CDF if the embedded line is Li . Substituting FC (c) from (3) and defining β π M − 2i = − β, 2 4 2 β π M − 2(i − 1) = ϕi − = − β, 2 4 2 ϕ′i = ϕi + ϕ′i−1 and di = σb|i σa|i P + (e|i) = di tan ϕ′i−1 − r|i 1 √ tan−1 π 1 − r2 |i where, − 2 2αi2 2 σ + 2σn2 , σb2|i = σ 2 + 2σn2 , (18) 1 + αi2 u 1 + αi2 u 2αi σ2 αi σu2 1+α2i u =√ 2 r|i = . σa|i σb|i (αi σu2 + (1 + αi2 )σn2 )(σu2 + (1 + αi2 )σn2 ) (19) σa2|i = ′ −1 di tan(−ϕi−1 ) 1 tan π + 1− + √ 1 − r|i2 Pe+ M 1 ∑ + = P (e|i), M i=1 − r|i di tan(ϕ′i ) − r|i 1 √ tan−1 π 1 − r|i2 di tan(−ϕ′i ) − r|i 1 √ , tan−1 π 1 − r2 (23) |i where σa|i , σb|i , and r|i are defined as in (18) and (19). Moreover, by defining Hi (ϕ) = di tan ϕ − r|i di tan(ϕ) + r|i 1 1 √ + tan−1 , tan−1 √ 2 π π 1 − r|i 1 − r|i2 we can simplify (23) as: P + (e|i) = 1 + Hi (ϕ′i−1 ) − Hi (ϕ′i ). (24) For i = 1 and i = M , we can see that: β β P + (e|1) = 1 − P (tan (−ϕ1 − ) < c < tan (ϕ1 + )|1) 2 2 = 1 − H1 (ϕ′1 ), (25) IV. P ERFORMANCE E VALUATION Here, we analytically study the error probability of the proposed watermarking scheme in the presence of AWGN. First, we compute Pe+ , the error probability when the slope of embedding line (p, q) is positive, given as: (22) , we have: |i These parameters are computed by substituting k as defined in (11) in definitions of variances of a, b and their correlation coefficient r. As shown in Appendix B, the decision boundary between each two coding line is their bisector. Considering this ML decision for all pairs of coding lines, we deduce that fC (c|i) is maximum if the angle of the received line with slope c falls between the bisectors of ith coding line and its two neighboring coding lines. Therefore, we obtain the decision boundaries as shown in Fig. 1(b) with dotted lines. It is notable that the deduced decoder is independent of β, the angle between two consecutive coding lines. β β ) < c < tan (ϕi − )|i) 2 2 P + (e|M ) = 1 − P (c > tan (ϕM − − P (c < tan (−ϕM + β )|M ) 2 β )|M ) = HM (ϕ′M −1 ), 2 (26) Substituting (24),(25), and (26) in (20),we have: = M M −1 1 ∑ M −1 1 ∑ + Hi (ϕ′i−1 ) − Hi (ϕ′i ) M M i=2 M i=1 = M −1 1 ∑ M −1 + [Hi+1 (ϕ′i ) − Hi (ϕ′i )], M M i=1 (20) where P + (e|i) represents the error probability when the embedding line is Li , the ith positive slope coding line. According to (16), for i = 2 · · · M − 1, the error probability Pe+ (27) Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. IMAGE WATERMARKING 5 (a) (b) (c) Fig. 2. Comparison between the theoretical error probability and experimental one for a Gaussian random variable with σu = 40. (a) one bit embedding (M = 2), (b) two bits embedding (M = 4), and (c) three bits embedding (M = 8). Fig. 3. Comparing the probability of error with DWR of 19 dB for different bit rates. With a similar argument we can see that the error probability for the negative slope embedding line is the same as Pe+ ; that is Pe− = Pe+ . Besides, it is easy to show that the positive and the negative embedding are equally probable in general; therefore, we can write the error probability of the decoder as Pe = 1 + 1 − P + Pe = Pe+ . 2 e 2 (28) Here, we obtained a closed form solution for the error probability of the decoder. In Fig. 2 we compared this theoretical error probability with the experimental case of Gaussian random variable with σu = 40 for different capacities of one bit (M = 2), two bits (M = 4), and three bits (M = 8). In this figure, the probability of error is shown versus various noise strengths, measured by the watermark-tonoise ratio (WNR), which is the ratio between the embedding distortion and the attacking noise distortion. As, we can see the theoretical and experimental results match perfectly. Fig. 4. Comparing the probability of error for the proposed method and the AQIM method [25] with DWRs of 19 dB and 22 dB. V. E XPERIMENTAL R ESULTS We performed several experiments to test the proposed algorithm and evaluate its performance against different attacks. A. AWGN attack on the artificial signal In the first experiment, to validate our method, we present the performance results for an artificial Gaussian signal under AWGN attack. We conducted the simulations for various bit rates and noise levels. The strength of the watermarking in terms of the document-to-watermark ratio (DWR) is fixed to 19dB and bit rates are set to one bit (M = 2), two bits (M = 4), and three bits (M = 8) per each four samples. The results versus various WNRs are given in Fig. 3. As expected, the performance degrades as M increases. However, we observe that the proposed scheme is practically applicable for M = 2, M = 4, and M = 8. Nevertheless, for further experiments we just use M = 2 and M = 4. In Fig. 4, we also compare the robustness for M = 2 with an AQIM based method [25] in similar situations for DWRs of Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. 6 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. ??, NO. ?, APR 2011 Fig. 5. Comparing the probability of error for the proposed method and the hyperbolic RDM scheme [28] under power-law attack. 19 dB and 22 dB. As we can see, the proposed method which employs an optimal decoder has lower error probability values even for very low WNRs. B. Power-law attack on the artificial signal Here, we show the robustness of the proposed watermarking algorithm against the power-law attack which can model the gamma correction attack (as a non-linear gain attack) that is a common process on image and video signals [28]. If we represent the watermarked signal with u′′ , the AWGN signal with n, and the attacked signal with z, the power-law attack is defined as z = λ(u′′ + n)γ , where λ is the gain factor and γ is the exponential parameter. Fig. 5 shows the performance of our sample projection approach against this attack for λ = 0.7 and γ = 1.2 along with the results obtained by [28] with different value of L in similar conditions. Here, L determines the memory of the system. As seen, we have high robustness against the powerlaw attack which is an important factor of our algorithm caused form the invariance of the proposed sample projection scheme to the gain attack. C. Tests on the image signal Then, we conduct several experiments to verify the performance of the proposed technique in the real application of image watermarking. Throughout our experiments, we use the Daubechies length-8 Symslet filters with two levels of decomposition to compute the 2-D DWT. The watermark data is embedded in the second level approximation coefficients of each block. The results are obtained by averaging over 100 runs with 100 different pseudorandom binary sequences as the watermark bits. For this study, we use four natural images of size 512×512: Plane, Pirate, Boat, and Bridge. The original test images and their watermarked versions using the proposed method with 16 × 16 block size and 128 bits message length are shown in Fig. 6. Original (top) and watermarked (bottom) test images; Left-right: Plane, Pirate, Boat, and Bridge. Fig. 6. Besides, β, the angle between two consecutive coding lines used in (11) is set to β = 40o for M = 2 and β = 20o for M = 4. These values are hand-optimized to achieve the highest robustness while keep the watermark almost imperceptible. As we can see, the watermark invisibility is perfectly satisfied. The mean peak-signal-to-noise-ratio (PSNR) of the watermarked images are 40.39dB, 40.44dB, 39.89, and 40.85, respectively. To quantitatively evaluate the imperceptibility of the watermarked images, we use the mean structural similarity index (MSSIM) [30] which is highly matched with the human visual system and effectively captures local errors in the spatial domain. The respective MSSIM values for four test images are 0.9983, 0.9984, 0.9982, and 0.9984 which confirms the imperceptibility of the proposed algorithm. Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. IMAGE WATERMARKING 7 TABLE III BER(%) R ESULTS OF EXTRACTED WATERMARK ATTACK . (a) 0 0.05 0.1 w 0.15 4.00 8.00 16.00 32.00 0.00 0.00 0.00 0.00 0.83 2.24 0.00 0.00 2.17 5.62 0.00 0.00 3.93 8.02 0.00 0.00 0.2 0.25 5.88 10.71 0.00 0.00 7.54 13.43 0.00 0.00 TABLE IV E FFECT OF BLOCK SIZE AND MESSAGE LENGTH ON THE WATERMARK ROBUSTNESS (BER (%)). (b) Fig. 7. AWGN attack for various noise variances, (a) for M = 2 (binary case) and (b) for M = 4 (2-bits). (a) Segment Size UNDER VARIABLE GAIN Blk. Size Bits MSSIM JPEG 20% AWGN 15 Gaus. 3×3 Crop 5% Scale 0.75 Rot. -5 8 8 8 8 16 16 16 32 32 128 256 512 1024 128 256 512 128 256 0.9995 0.9989 0.9980 0.9961 0.9983 0.9965 0.9932 0.9906 0.9850 10.92 11.38 13.60 16.93 1.92 2.58 4.96 1.09 6.27 9.14 9.89 12.22 14.82 4.27 4.81 7.99 1.70 7.13 10.61 11.25 12.67 12.71 1.18 2.31 2.33 0.12 0.48 4.14 4.38 4.87 5.24 4.29 4.71 5.89 9.49 11.35 3.81 3.57 4.36 4.09 0.19 0.12 0.07 0.00 0.00 7.95 8.07 8.54 9.09 2.60 2.75 3.56 4.90 5.44 (b) Fig. 8. JPEG compression attack for various quality factors, (a) for M = 2 (binary case) and (b) for M = 4 (2-bits). Image Median Filtering 3×3 M =2 Plane Pirate Boat Bridge 1.50 4.05 3.34 3.80 0.38 1.58 1.74 1.00 1.15 2.38 2.43 2.66 1.15 2.82 2.84 2.86 M =4 TABLE I BER(%) R ESULTS OF EXTRACTED WATERMARK UNDER MEDIAN AND G AUSSIAN FILTERING ATTACKS . Gaussian Filtering 3×3 5×5 7×7 Plane Pirate Boat Bridge 6.57 7.38 7.02 6.93 4.61 3.95 6.05 4.59 5.52 5.66 9.22 6.87 5.79 5.70 9.38 7.16 1) AWGN attack: As the first attack, we investigate the effect of AWGN to the proposed watermarking scheme. Fig. 7 shows the bit error rate (BER) of the proposed method for various images versus different noise powers. As we see, the method has a great resistance against noise attack for both M = 2 and M = 4. 2) JPEG attack: Secondly, the proposed technique is tested against JPEG compression with different quality factors. As demonstrated in Fig 8, since low frequency components of the image blocks are used for watermarking, the proposed method is highly robust against JPEG with different quality factor up to 10%. 3) Filtering attack: Table I shows the BER results for median filtering, and Gaussian low-pass filtering attacks for different test images. It can be seen that the proposed scheme is highly robust against these attacks. 4) Editing attack: In Table II, we also report the performance of our algorithm under several image editing attacks such as cropping, scaling, and rotation. As we can see in this table, the proposed scheme remains highly robust against these attacks as well. 5) Gain attack: To show the robustness of the proposed algorithm to constant and variable gain attack, we conduct an experiment on the four test images. To model the variable gain attack, we segment each image to small non-overlapping blocks and multiply all the pixels in each block by the gain factor of 1 + wδ where δ is a random variable uniformly distributed between 0 and 1, and w is a constant determining the strength of the attack . Thus, over each block, the gain factor remains constant, while for each block we use different gain factors. This modeling can realize the variable gain attack and fading attack happens through scanning process. Results for different segmenting block sizes and strength factors ws are shown in Table III ( Results are averaged over the four test images for M = 2). As we see, for block sizes of 16 and 32 our algorithm is invariant to the gain attack. This is because our watermarking approach embeds information in blocks of size 16; hence, the gain will be constant in each embedding block, and thud in the approximation band. That is, duo to the invariance of the sample projection scheme to the constant gain factor, we will have 0% BER for segmentation of 16 and 32. For segments of size 8 and 4 we will have variable gain factors in each embedding block. However, we can see that even for such variable gains, our algorithm shows high robustness. Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. 8 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. ??, NO. ?, APR 2011 TABLE II BER(%) R ESULTS OF EXTRACTED WATERMARK UNDER CROPPING , SCALING , AND ROTATION M =2 Image Plane Pirate Boat Bridge Cropping 5% 10% 3.95 4.42 3.57 5.23 5.71 5.71 6.20 11.34 ATTACKS . M =4 0.5 Scaling 0.75 1.5 -10 0.38 0.70 0.72 1.42 0.00 0.38 0.00 0.37 0.00 0.00 0.00 0.00 2.23 5.06 2.72 6.67 Rotation -5 5 1.45 3.77 1.54 3.64 0.69 2.73 0.35 4.75 10 1.09 3.11 1.12 5.87 Cropping 5% 10% 4.00 5.57 4.68 4.80 5.97 5.72 6.84 10.94 0.5 Scaling 0.75 1.5 -10 2.54 4.08 4.25 4.74 0.00 0.60 0.48 0.76 0.00 0.81 0.00 0.24 5.16 6.30 7.53 5.50 Rotation -5 5 3.83 5.31 4.53 4.70 10 2.37 4.23 2.58 3.37 2.26 4.45 3.20 5.62 TABLE V C OMPARISON BETWEEN OUR WATERMARKING METHOD AND [24] : BER (%) UNDER AWGN ATTACK . σn 4 Method [24] Proposed 1 2 3 1.00 0.00 2.00 0.00 3.00 0.12 4.00 0.12 5 6 7 15.00 0.43 25.00 0.85 44.00 1.05 TABLE VI C OMPARISON BETWEEN OUR WATERMARKING METHOD AND WANG ’ S METHOD [31] : BER (%) UNDER M EDIAN F ILTERING ATTACK WITH WINDOW SIZE 3 × 3. Method Barbara Wang [31] Proposed Fig. 9. 24.95 7.00 Image Baboon Peppers 31.65 9.88 29.35 3.25 Goldhill 25.60 1.65 Capacity of the proposed method for BER less than 5%. 6) Block size and bit rate effect: In this experiment, we study the effect of changing the block size and message length on the performance of the proposed algorithm. To this aim, we vary the block size from 8 to 32, and the number of blocks engaged, that is the message length, from 128 to 1024. The robustness for M = 2, averaged over the four test images, against set of attacks as well as the MSSIM value, measure of imperceptibility, is demonstrated in Table IV. Our experiments on various block sizes and watermark message lengths indicated that while other block size and bit rates also can be picked based on the application, the case of 16 × 16 block size and 128 bit message length is the best compromise between imperceptibility, capacity, and robustness. 7) Capacity: To show the capacity of the proposed scheme, a graph of capacity versus noise standard deviation is shown in Fig. 9. In this graph, we have considered a fixed BER of 5% as a typical sufficiently small error rate and investigated the maximum message length in bits for 16 × 16 block size and M = 2 which obtains this error rate. The average message length over four test images as well as the average MSSIM value for each case is shown in Fig. 9. As shown in this figure, even for heavy noise powers of σn = 20 , the capacity of the proposed method in obtaining the BER of less than 5% is near 140 bits. 8) Comparison With Other Watermarking Schemes: Finally, we compare our watermarking algorithm for M = 2 with two of the recent blind watermarking techniques, [24] and [31], in the same situation, for AWGN and median filtering attacks. The simulation results are shown in Tables V, VI. As shown in these tables, the robustness of our method is considerably better than these two techniques. VI. C ONCLUSION In this article, we presented a novel blind watermarking approach with the optimal decoder. Watermark embedding is performed by multiplication of M specific matrices to the vector of samples of size four. These matrices translate and project the vector of samples on the predefined coding lines depending on the message symbol. Assuming the host samples to be i.i.d Gaussian, which is often valid for approximation coefficients of image blocks [15], we obtained a closed form PDF of noisy watermarked samples. Having this distribution function, we designed an optimum ML decoder. We analytically studied and verified the error probability of the proposed decoder in a noisy environment. The proposed algorithm is applied to image signals by using four approximation coefficients of the image blocks which may be selected according to a secret key. In addition to be invariant to the volumetric distortions, several simulations showed that the proposed algorithm is highly robust against common watermarking attacks such as AWGN, compression, and filtering. However, the algorithm is sensitive to the collusion attack. The future work may be proposing a solution to this problem in order to make the Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. IMAGE WATERMARKING 9 algorithm robust to the this attack as well. where, A PPENDIX A C OMPUTING THE P ROBABILISTIC C HARACTERISTICS OF THE D ECISION VARIABLE To compute the PDF and the CDF of the variable c = ab , for correlated Gaussian variables a ∼ N (0, σa2 ) and b ∼ N (0, σb2 ) with correlation coefficient r, consider their joint distribution, given as: fab (a, b) = 1 √ 2 2 1 a 2rab b − 2(1−r 2 ) ( σ 2 − σa σ + 2 ) 2πσa σb 1 − r2 e a b σ b . (29) a b Using this distribution function, the CDF of c = can be computed as: } {a ≤c FC (c) = P {b } { } = P a ≤ bc, b > 0 + P a ≥ bc, b < 0 ∫ ∞ ∫ bc = fab (a, b) da db b=0 a=−∞ ∫ 0 ∫ ∞ + fab (a, b) da db, (30) b=−∞ a=bc and its probability density function can be defined as: ∫ ∞ ′ fC (c) = FC (c) = |b|fab (bc, b) db. (31) −∞ Substituting (29) in (31) and considering the fact that fab (a, b) is an even function with respect to a and b, we have: ∫ ∞ 2 − b2 1 σ02 √ √ fC (c) = , be 2σ0 db = 2πσa σb 1 − r2 0 πσa σb 1 − r2 (32) where 1 − r2 σ02 = c 2 . ( σa ) − σ2rc + σ12 a σb b As a consequence, fC (c) = √ 1 σa σb 1 − r 2 . a 2 2 2 π σb2 (c − rσ σb ) + σa (1 − r ) In addition, FC (c) can be computed as: ∫ c 1 1 σb c − rσa FC (c) = fC (t) dt = + tan−1 √ . 2 π σa 1 − r 2 −∞ (33) (34) A PPENDIX B C OMPUTING THE D ECISION B OUNDARIES OF THE ML D ECODER For each two coding lines Li and Lj we have: fC (c|i) ≷ij fC (c|j). (35) Substituting (17) in (35) and after some simplifications, we obtain: m2 c2 + m1 c + m0 ≷ij 0, (36) m2 = m1 = m0 = σb|j √ σb|i √ 2 1 − r|i2 − 1 − r|j σa|j σa|i √ √ 2 − 2r 2r|i 1 − r|j 1 − r|i2 |j σa|j √ σa|i √ 2. 1 − r|i2 − 1 − r|j σb|j σb|i (37) To find the roots for this quadratic equation, we show these coefficients using ϕi = arctan(αi ), the angle of the coding σ2 line Li . Thus, by defining γ 2 = σu2 , we can write the n parameters in (18) and (19) as: σa2|i = 2σn2 (γ 2 sin2 (ϕi ) + 1), r|i = σb2|i = 2σn2 (γ 2 cos2 (ϕi ) + 1) sin(ϕi ) cos(ϕi )γ 2 sin(2ϕi )σu2 =√ . σa|i σb|i (γ 2 sin2 (ϕi ) + 1)(γ 2 cos2 (ϕi ) + 1) (38) Then, it is easy to see that: √ √ 2σn2 γ 2 + 1 2 1 − r|i = . σa|i σb|i (39) Using these equations, we can rewrite m2 in (37) as: √ 2σn2 γ 2 + 1 m2 = (σb2|j − σb2|i ) σa|i σb|i σa|i σb|j [ ] 2 = R cos (ϕj ) − cos2 (ϕi ) , = R sin(ϕi + ϕj ) sin(ϕi − ϕj ), (40) √ 2 4 2 4σ γ γ +1 where R = σa nσb σa σb . Similarly m0 can be computed as |i |i |i |j m0 = −m2 , and m1 can be rewritten as: [ ] m1 = R sin(2ϕi ) − sin(2ϕj ) = 2R cos(ϕi + ϕj ) sin(ϕi − ϕj ). (41) Therefore, the roots of (36) can be found as: c = tan( ϕi + ϕj ) or 2 c=− 1 ϕ +ϕ tan( i 2 j ) , (42) which means the decision boundary between each two coding lines Li and Lj . is either their bisector or its perpendicular line. Since we first realize the sign of c and then detect its corresponding coded word, the only valid solution will be the bisector line. R EFERENCES [1] J. Seitz, Digital Watermarking For Digital Media. Arlington, VA, USA: Information Resources Press, 2005. [2] C.-S. Lu, Multimedia Security: Steganography and Digital Watermarking Techniques for Protection of Intellectual Property. Hershey, PA, USA: IGI Publishing, 2004. [3] G. Langelaar, I. Setyawan, and R. Lagendijk, “Watermarking digital image and video data. a state-of-the-art overview,” IEEE Signal Process. Mag., vol. 17, no. 5, pp. 20 –46, Sep. 2000. [4] I. Cox, J. Kilian, F. Leighton, and T. Shamoon, “Secure spread spectrum watermarking for multimedia,” IEEE Trans. Image Process., vol. 6, no. 12, pp. 1673 –1687, Dec. 1997. [5] J. Hernandez, M. Amado, and F. Perez-Gonzalez, “Dct-domain watermarking techniques for still images: detector performance analysis and a new structure,” IEEE Trans. Image Process., vol. 9, no. 1, pp. 55 –68, Jan. 2000. Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected]. This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. 10 IEEE TRANSACTIONS ON INFORMATION FORENSICS AND SECURITY, VOL. ??, NO. ?, APR 2011 [6] H. Altun, A. Orsdemir, G. Sharma, and M. Bocko, “Optimal spread spectrum watermark embedding via a multistep feasibility formulation,” IEEE Trans. Image Process., vol. 18, no. 2, pp. 371 –387, Feb. 2009. [7] Q. Cheng and T. Huang, “An additive approach to transform-domain information hiding and optimum detection structure,” IEEE Trans. Multimedia, vol. 3, no. 3, pp. 273 –284, Sep. 2001. [8] M. Kutter and S. Winkler, “A vision-based masking model for spreadspectrum image watermarking,” IEEE Trans. Image Process., vol. 11, no. 1, pp. 16 –25, Jan. 2002. [9] P. Moulin and A. Ivanovic, “The zero-rate spread-spectrum watermarking game,” IEEE Trans. Signal Process., vol. 51, no. 4, pp. 1098 – 1117, Apr. 2003. [10] S. Maity and S. Maity, “Multistage spread spectrum watermark detection technique using fuzzy logic,” IEEE Signal Process. Lett., vol. 16, no. 4, pp. 245 –248, Apr. 2009. [11] M. Barni, F. Bartolini, A. De Rosa, and A. Piva, “A new decoder for the optimum recovery of nonadditive watermarks,” IEEE Trans. Image Process., vol. 10, no. 5, pp. 755 –766, May. 2001. [12] ——, “Optimum decoding and detection of multiplicative watermarks,” IEEE Trans. Signal Process., vol. 51, no. 4, pp. 1118 – 1123, Apr. 2003. [13] Q. Cheng and T. Huang, “Robust optimum detection of transform domain multiplicative watermarks,” IEEE Trans. Signal Process., vol. 51, no. 4, pp. 906 – 924, Apr. 2003. [14] J. Wang, G. Liu, Y. Dai, J. Sun, Z. Wang, and S. Lian, “Locally optimum detection for barni’s multiplicative watermarking in dwt domain,” Signal Process., vol. 88, no. 1, pp. 117–130, 2008. [15] M. A. Akhaee, S. M. E. Sahraeian, B. Sankur, and F. Marvasti, “Robust scaling-based image watermarking using maximum-likelihood decoder with optimum strength factor,” IEEE Trans. Multimedia, vol. 11, no. 5, pp. 822 –833, Aug. 2009. [16] M. A. Akhaee, S. M. E. Sahraeian, and F. Marvasti, “Contourlet-based image watermarking using optimum detector in a noisy environment,” IEEE Trans. Image Process., vol. 19, no. 4, pp. 967 –980, Apr. 2010. [17] S. M. E. Sahraeian, M. A. Akhaee, and F. Marvasti, “Distribution independent blind watermarking,” in Proc. IEEE Int. Conf. Image Processing, Cairo, Egypt, Nov. 2009, pp. 125 –128. [18] C.-W. Tang and H.-M. Hang, “A feature-based robust digital image watermarking scheme,” IEEE Trans. Signal Process., vol. 51, no. 4, pp. 950 – 959, Apr. 2003. [19] D. Zheng, J. Zhao, and A. El Saddik, “Rst-invariant digital image watermarking based on log-polar mapping and phase correlation,” IEEE Trans. Circuits Syst. Video Technol., vol. 13, no. 8, pp. 753 – 765, Aug. 2003. [20] B. Chen and G. Wornell, “Quantization index modulation: a class of provably good methods for digital watermarking and information embedding,” IEEE Trans. Inf. Theory, vol. 47, no. 4, pp. 1423 –1443, May. 2001. [21] J. Eggers, R. Buml, and B. Girod, “Estimation of amplitude modifications before scs watermark detection,” in Proc. SPIE: Security Watermarking Multimedia Contents IV, vol. 46, San Jose, Ca, USA, Jan. 2002, pp. 387–398. [22] J. H. Conway, N. J. A. Sloane, and E. Bannai, Sphere-packings, lattices, and groups. New York, NY, USA: Springer-Verlag New York, Inc., 1998. [23] M. Miller, G. Doerr, and I. Cox, “Applying informed coding and embedding to design a robust high-capacity watermark,” IEEE Trans. Image Process., vol. 13, no. 6, pp. 792 –807, Hun. 2004. [24] C. Chen and X. Wu, “An angle qim watermarking algorithm based on watson perceptual model,” in Proc. Int. Conf. Image and Graphics Processing (ICIG07), Chengdu, Sichuan, China, 22-24 2007, pp. 324 –328. [25] F. Ourique, V. Licks, R. Jordan, and F. Perez-Gonzalez, “Angle qim: a novel watermark embedding scheme robust against amplitude scaling distortions,” in Proc. IEEE Int. Conf. Acoustics, Speech, and Signal Processing, vol. 2, Philadelphia, PA, USA, Mar. 2005, pp. 797 – 800. [26] F. Perez-Gonzalez, C. Mosquera, M. Barni, and A. Abrardo, “Rational dither modulation: a high-rate data-hiding method invariant to gain attacks,” IEEE Trans. Signal Process., vol. 53, no. 10, pp. 3960 – 3975, Oct. 2005. [27] F. Perez-Gonzalez and C. Mosquera, “Quantization-based data hiding robust to linear-time-invariant filtering,” IEEE Trans. Inf. Forensics Security, vol. 3, no. 2, pp. 137 –152, Jun. 2008. [28] P. Guccione and M. Scagliola, “Hyperbolic rdm for nonlinear valumetric distortions,” IEEE Trans. Inf. Forensics Security, vol. 4, no. 1, pp. 25 –35, Mar. 2009. [29] M. A. Akhaee, A. Amini, G. Ghorbani, and F. Marvasti, “A solution to gain attack on watermarking systems: Logarithmic homogeneous rational dither modulation,” in Proc. IEEE Int. Conf. Audio, Speech, and Signal Process., Dallas, TX, USA, May. 2010, pp. 1050 –1053. [30] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image quality assessment: from error visibility to structural similarity,” IEEE Trans Image Process, vol. 13, pp. 600–612, Apr 2004. [31] Y. Wang, J. Doherty, and R. Van Dyck, “A wavelet-based watermarking algorithm for ownership verification of digital images,” IEEE Trans. Image Process., vol. 11, no. 2, pp. 77 –88, Feb. 2002. Mohammad Ali Akhaee (S’07–M’10) was born in Tehran, Iran, in 1982. He received the B.S. degree in both electronic and communication engineering from Amirkabir University of Technology and the M.S. degree from Sharif University of Technology, Tehran, Iran, in 2003 and 2005, respectively. He received his PhD degree from Sharif University of Technology in 2009. He has awarded the Endeavour research fellowship from Australia in 2010. He is now a post-doc researcher at the Computing and Audio Research Laboratory at the University of Sydney. His research interest includes multimedia security, statistical signal processing, acoustic, and image processing. Sayed Mohammad Ebrahim Sahraeian (S’08) was born in Shiraz, Iran, in 1983. He received the B.S. and M.S. degrees in electrical engineering from Sharif University of Technology, Tehran, Iran, in 2005 and 2008, respectively. He is currently pursuing the Ph.D. degree in electrical and computer engineering at Texas A&M University, College Station. His research interests include image and multidimensional signal processing, genomic signal processing, bioinformatics, and computational biology. Craig Jin (M’92) received the M.S. degree in applied physics from the California Institute of Technology, Pasadena, in 1991 and the Ph.D. degree in electrical engineering from the University of Sydney, Sydney, Australia, in 2001. He is a Senior Lecturer in the School of Electrical and Information Engineering, University of Sydney and also a Queen Elizabeth II Fellow. He is a Director of the Computing and Audio Research Laboratory at the University of Sydney and also a co-founder of three start-up companies. His research focuses on spatial audio and neuromorphic engineering and he is the author or co-author of more than 50 journal or conference papers in these areas and holds six patents. Dr. Jin has received national recognition in Australia for his invention of a spatial hearing aid. Copyright (c) 2011 IEEE. Personal use is permitted. For any other purposes, Permission must be obtained from the IEEE by emailing [email protected].
© Copyright 2024