Image Watermarking Scheme With Unequal

JOURNAL OF MULTIMEDIA, VOL. 5, NO. 5, OCTOBER 2010

427

Image Watermarking Scheme With Unequal Protection Capability Based on Error Correcting Codes Chuan Qin1, Qian Mao1, and Xinpeng Zhang2 1

School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, Shanghai 200093, China 2 School of Communication and Information Engineering, Shanghai University, Shanghai 200072, China E-mail: [email protected], [email protected], [email protected]

Abstract—We propose a novel image watermarking scheme capable of providing unequal protection levels to different regions. In the scheme, we firstly construct the unequal error protection (UEP) code by designing an unequal coding space using algebra means, then we divide the image into high protection region, low protection region and non-protection region, and use the UEP code to encode the bits generated by high and low protection regions simultaneously and form the watermark bits to embed in the non-protection region. On the receiver side, the extracted bits with UEP capability can restore the high protection region more correctly than the low protection region after the efficient decoding algorithm is applied. Index Terms―image watermarking, unequal protection, error correcting codes, image restoration

I.

INTRODUCTION

With the rapid development of Internet and multimedia technology, copyright protection becomes more and more important. Digital watermarking which is one kind of information hiding technology can be used as a tool of protecting digital contents, and arises much interest of many researchers [1-2]. While robust watermarking can be used for ownership verification, fragile watermarking are intended for checking integrity and authenticity of digital contents. We mainly discuss the scenarios of image authentication and recovery in this paper. Some fragile watermarking methods are designed to locate the modified or tampered regions. Many schemes usually divide a host image into small blocks and embed the watermark into each block [3-5]. The embedded data may be a hash of the principal content of each cover block. If the image has been modified, the image content and the watermark corresponding to the tampered blocks cannot be matched so that the tampered blocks are detected. Moreover, some watermarking methods can also recover the tampered areas after the tampering detection. Two image recovery approaches based on block-wise fragile watermarking have been proposed in [6]. With the first method, the DCT coefficients of every block sized 8×8 are quantized into 64 or 128 bits, which are used to replace one or two least significant bits of another block. In the second method, a low color depth version of the original image generated by reducing gray levels is cyclic-shifted and embedded into the

© 2010 ACADEMY PUBLISHER doi:10.4304/jmm.5.5.427-433

pixel differences. After the malicious modification on a watermarked image is localized, the quantized DCT coefficients and the low color depth data extracted from reserved regions can be exploited to recover the principal content of tampered areas. But all this kind of block-wise methods can only detect and restore tampered blocks, but not the tampered pixels. Zhang, et al., proposed a statistical fragile watermarking method which can locate individual tampered areas in the pixel-wise precision [7]. This method embeds a set of tailormade authentication data into a host image and introduces a statistical mechanism for image authentication. By estimating the modification strength, two different distributions corresponding to tampered and original pixels can be used to exactly locate the tampered pixels. Zhang, et al., improved the method in [7] to integrate the recovery capability in their fragile watermarking scheme [8]. In their scheme, a tailor-made watermark consisting of reference-bits and check-bits is embedded into the host image using a lossless data hiding method. On the receiver side, by comparing the extracted and calculated check-bits, the tampered image blocks can be identified. Then the reliable reference-bits extracted from other blocks are used to exactly reconstruct the original image. When the tampered area is not too extensive, this method can restore the original image information with no error. Feng, et al., proposed an unsupervised region-level image authentication method named Bayesian structural content abstraction (BaSCA) [9]. They model image structural content using the net-structured Markov pixon random field and derive the size-controllable BaSCA signature, which is capable of tolerating a wide range of noncontent changing operations robustly and can detect contentchanging operations sensitively. But the mostly reported watermarking methods treat the importance of different image regions as the same, and can’t give the unequal protection capability for the different regions. However in practice, different regions or contents of the image may have different importance for users. When the image suffers tampering, the regions or the contents which the user shows more interests to should have higher protection level, and need to be restored more accurately. In this paper, we propose a novel watermarking scheme which can provide unequal protection levels for different regions of the image.

428


In the following, Section II describes our watermarking scheme based on unequal error protection (UEP) codes. Section III presents the experimental results and discussion, and Section IV concludes the paper. II.

WATERMARKING WITH UNEQUAL ERROR PROTECTION CAPABILITY

Different regions of the image often have unequal importance, and attract different degrees of attention and interest for users. So we indicate that different regions should have unequal protection levels if the image suffers tampering and modification. In this paper, we use UEP codes to realize the function of unequal protection for the image. The UEP code algorithm in our method encodes the watermarking bits generated by different image regions simultaneously, but can give unequal protection levels to them. The present scheme is concise because only one encoder and decoder is needed, and due to the improved decoding algorithm, the UEP code doesn’t decrease the coding rate. A. Unequal Error Protection Codes There are two sorts of UEP codes: message-wise UEP codes and bit-wise UEP codes. The former is used in the situations of providing better protection to a subset of messages. It is realized by multiple coding schemes, so the structures of the encoder and decoder are more complicated. The latter was first proposed by Masnick in [12], it is often used to provide better protection to one subset of bits in a coding block and ensure that the bits of higher protection level have less error probability than the bits with lower protection level [10-12]. The bit-wise method constructs unequal distances for information bits in one coding block by algebraic means and assigns them with different error correcting capabilities in once coding process without changing the coding rate. So in this paper, we adopt the bitwise UEP code because it uses only one encoder and decoder to realize the unequal error protection in bit-level. Suppose that the generator polynomial g(x) of the binary cyclic UEP code C (n, k) is a direct sum of v ideals , , …, defined in the ring GF(x)/, and the separation vector is U = (u1, u2, …, uk).

where ⎣x⎦ denotes the integer not larger than x. It means the whole codeword will be decoded correctly if there are ε errors in a codeword and ε is not larger than the minimal Hamming distance t. If ε is larger than t, the whole codeword may not be decoded correctly, but all the information bits whose error protection capabilities fi aren’t smaller than ε will be decoded correctly. To obtain unequal error protection capability, we should construct the optimal generation matrix:

G opti = [G 1 G 2 L G v ]T

(4)

where Gi (i=1, …, v) is the generation matrix of the ideal [13]. The codeword c can be obtained after the encoding:

c = m G opti

(5)

The twice-decoding algorithm is utilized in the decoding process to provide unequal error protection capability for information bits m. Suppose there are two levels of error correcting capabilities: f1 and f2 (f1>f2). If the number of error bits ε in a codeword is not larger than f2, the decoder will process the codeword only using the general decoding algorithm.

Si + σ 1Si−1 + L + σ t Si−t = 0 i = 1,2, ..., t

where Si is the syndrome of code C(n, k) and σ is error location. If ε > f2, another step is needed in the decoding procedure, see Equation (7). Suppose that m’ is the first k1 information bits of m whose error correcting capability is f1. And then we can obtain the subcode c’’ (containing the error vector e) whose error correcting capability is f2.

c '' ⊕ e = r ⊕ m 'G '

(7)

0

10

-1

m = (m0 , m1 , L , m k −1 ) ∈ GF , and mi ≠ 0

10

(1)

In Equation (1), W[·] means the Hamming weight of the code, m(x) is the information polynomial. Then the error protection vector F of the cyclic UEP code is (f1, f2, …, fk).

⎢ u − 1⎥ fi = ⎢ i ⎥ ⎣ 2 ⎦

i = 1, 2, ..., k

t = min ( f1 , f 2 , ..., f k )

© 2010 ACADEMY PUBLISHER

-2

10 BER

u i = min {W [m ( x ) g ( x ) ]}, subject to

(2)

-3

10

m0~m1 -4

10

m2~m6

-5

10

-3

10

(3)

(6)

-2

10 Channel Error Rate

-1

10

Figure 1. BER performance of UEP code


High protection region Original Image

Region Division

Low protection region

429

High Level Bits

Low Level Bits

Unequal Error Protection Encoding

Interlace

Watermarking Bits

Non-protection region

Embedding Procedure

Watermarked Image

Figure 2. UEP watermark embedding procedure where G’ is the generator matrix of the subcode constructed by m’, and r is the received vector. If the syndromes of c’’⊕e show that there are more than f2 errors in the subcode and the syndromes are self-consistent, the hypothesis of m’ is correct. Otherwise the decoder changes the hypothesis and checks the syndromes of c’’⊕e again until the correct m’ is found. Figure 1 shows the bit error rate (BER) performance of the (15,7) UEP code, and its error protection vector F is (2,2,1,1,1,1,1). This code provides 2-bit error correction capability to the first two information bits (m0~m1) and 1-bit error correction capability to the last five information bits (m2~m6). We can fnd in Figure 1 that the BER performance of the first two bits is better than that of the last five bits. B. Watermarking Scheme Using UEP Codes We all know that the observer may pay more attention to the region which he is aware of important or interesting in the image. Usually, the foreground object is the most important portion of the image, and the neighborhood of the foreground object maybe the less important portion, while the background is often seemed as the least important region which attracts little attention. And we think the protection level should be direct proportion to the importance of the image region. The region which the observer shows more interest maybe more important and should be given higher level protection. According to the above analysis, we first separate the image into three kinds of regions: high protection region (HPR), low protection region (LPR) and non-protection region (NPR), and then the pixels in HPR and LPR are coded by our UEP encoding algorithm. After that, the encoded bits are embedded into the NPR of the image by one typical embedding algorithm for robust watermarking with one secret key. The embedding locations of watermarking bits in NPR are decided by the secret key. The entire procedure of watermark embedding is sketched in Figure 2. The mask image set by user which labels the locations of three kinds of region should be sent to the receiver side for watermarking extraction and restoration. The mask image can also be regarded as another secret key, which can be securely transmitted by the key exchange protocol, such as DiffieHellman key exchange. Through the UEP encoding and decoding, our method can provide protection to the bits generated by the two kinds


of image region. Due to the function of unequal error protection of the method, the high level bits can be restored more correctly compared with the low level bits after the image suffers modification and tamper. One of the key steps in this method is how to rearrange the two kinds of bits by interlacing before inputting them to the encoder. The details of this process are expressed in the following. Denote the number of the pixels in the HPR which is masked by the user as p. We know every gray pixel can be expressed into 8 bits, so the generated high level bits have 8p bits: h={h1, h2, …, h8p}. In the same manner, the low level bits generated by LPR can be formed as l={l1, l2, …, l8q}, where q is the number of the pixels in the LPR. Suppose we choose a (n, k) UEP code in the method, which provides high protection for the first k1 information bits and low protection for the remaining k2 bits. k is the length of information vector while n is the length of encoded vector, and k1+ k2 ≡ k. Because the UEP code is one kind of linear block code, it operates one information vector during one encoding and decoding process. We first divide high level bits and low level bits generated by HPR and LPR into groups. Every group has k bits in which there are k1 bits belonged to high level bits and k2 bits belonged to low levels bits. If the number of two kinds of bits don’t satisfy Equation (8), we can pad or cut 0 or 1 with low level bits to meet the relation.

k1 p = k2 q

(8)

By this way in the k bits information vectors of UEP code, the k1 bits generated by HPR has higher protection level, while the k2 bits generated by LPR has lower protection level. After inputting all the groups of k bits information vectors into UEP encoder, we acquire the groups of n bits encoded vectors. Then the encoded bits with UEP capability can be regarded as watermark to embed in the NPR. The reason of embedding the watermark bits into NPR is that this region is usually unimportant visually and suffers less modification. If the area of HPR and NPR are too large, the sequence of generated information bits might be too long to embed into the remaining NPR. So before embedding, the estimation process will be conducted to decide if the NPR is

430

large enough to hide the embedding bits for the chosen embedding algorithm. On the receiver side, the watermark extracting is the inverse procedure of the embedding. The watermark bits are firstly extracted from NPR according to the mask image and secret key for watermark embedding, and then after decoding and de-interlacing the extracted bits can be used to restore the HPR and LPR from its tampered version. C. Security Analysis We also give the security analysis of the proposed watermarking scheme. Denote the region belonged to HPR or LPR as protection region (PR). There are mainly three scenarios for the attack stated in the following: 1) Only PR is tampered. If the attacker only tampers the PR of watermarked image, the watermarking bits can still be correctly extracted from NPR and used for PR restoration. Since all the bits are hidden in the NPR, the tampering will not affect the watermarking extraction. Where the content in PR doesn’t match the extracted bits can be detected as tempered region. 2) Only NPR is tampered. The proposed method doesn’t protect NPR of the image. We should note that the receiver doesn’t know whether the PR is tampered in this situation. The embedded bits hidden in NPR may be destroyed when the error due to severe tampering exceeds the errorcorrecting capability of the present UEP algorithm. But if these destroyed bits are used to “restore” the PR, the meaningless contents will be generated. The user will find that restoration process is unnecessary in this scenario or ask the sender to resend the image directly. 3) Both PR and NPR are tampered. If both these two regions are severely tampered by the attacker, the image will lose its use value. The receiver can ask the sender to resend


the image. If the attacker tries to carefully modify the PR and NPR in such manner: he first modifies the PR, and embeds the bits generated by the modified PR into NPR to form a tampered image, but the attacker still can’t achieve this kind of attack, because he doesn’t know the exact mask image and the secret key for embedding and extracting. The receiver needn’t know which kind of attack is occurred, when he or she starts to restore the images. No matter the attacker tampers the PR or NPR, the receiver still implements the extraction and restoration process. The only difference is that the receiver should ask the sender to resend the image when the restored result of PR is meaningless due to the severe tampering in NPR. Besides the malicious tampering, images will also be subject to noise interference during transmission in poor channel. The UEP codes algorithm can conduct error correction for the information bits embedded in the image. Based on the above analysis, the security of our watermarking scheme completely depends on the secret key, which satisfies the security requirement in the sense of cryptography. III.

RESULTS AND COMPARISONS

Experiments were carried out on a group of color and gray images with different sizes. For color images, the embedding and restoration are done on the R, G, and B channels respectively. The obtained components are then combined to give the final results. In the experiments, we choose (15, 13) UEP code and k1=1, k2=12. Encoding only introduces 2 bits redundancy. So in every information vector, the first bit belonged to HPR owns higher protection level while the remaining 12 bits belonged to LPR have lower protection level.

Figure 3. Restoration results of proposed UEP method. The first column shows the input original image and the mask image (Black: HPR, Gray: LPR, White: NPR) respectively. The 2nd through 4th images in the first row are the watermarked images with tampering and noises for different intensities (0.5%, 1.0%, and 1.5%), the 2nd through 4th images in the second row are the corresponding restoration results.



431

TABLE I. Comparison of PSNR (dB) Noise Intensity

PSNR in PR

PSNR in HPR

PSNR in LPR

0.5%

43.1

93.1

43.0

1.0%

39.3

62.1

39.1

1.5%

37.5

52.9

37.4

TABLE II. Comparison of UEP and EEP based methods Image

PSNR in PR (dB)

BER in HPR

BER in LPR

EEP

UEP

EEP

UEP

EEP

UEP

Bird

41.3

42.3

0.0125

0.0004

0.0142

0.0142

Flower

39.6

40.3

0.0143

0.0004

0.0158

0.0158

Plane

41.3

42.8

0.0244

0.0027

0.0304

0.0304

Coin

41.2

44.3

0.0247

0.0017

0.0301

0.0300

Figure 3 show some restoration results of proposed method. The first row shows the original image and the tampered images embedded watermarking. These tampered images also suffer noises with different intensities to show the error-correcting capability of the method. Salt and pepper noise is chosen in the experiment and other types of noise can also be applied, such as Gaussian noises. The noise intensities in the test are: 0.5%, 1.0%, and 1.5% respectively, which means the polluted percentage of pixels. The mask image representing protection levels is given in the first one of the second row. The black and gray areas in the mask image show HPR and LPR separately, and the white area is NPR. The 2nd through 4th images in the second row are the corresponding restoration results. We can find that HPRs of the damaged images are restored more correctly than LPRs by UEP based method. The PSNR evaluation is given in Table I. The PSNR values of restored HPRs are always higher than LPRs for different noise intensities. We then compare present UEP based method with the equal error protection (EEP) method whose decoder is only the general error correcting algorithm without UEP capability. Some images are given in Figure 4. The original input images and mask images are shown in the first two columns. The 3rd column is the tampered images with noises. The 4th column presents the restoration results of the method using EEP codes, while the results of our UEP based method are given in the last column. The performance of PSNR and bit error rate (BER) is also compared in the experiment. We can see from Table II that, our method has higher PSNR values in restored area and lower BER in the high protection region of the image, and since the UEP code has the same protection capability with EEP code in low protection region, we acquire the approximate BER in that region. By both subjective observation and objective evaluations with respect to the originals, we conclude that the proposed


UEP based method gives satisfactory output. Compared with the EEP based method, it produces better restoration quality and performance in terms of PSNR and BER. IV.

CONCLUSIONS

Observers are usually interested in some specified regions in the image, and don’t pay much attention to some unimportant regions. Different image regions should have unequal protection levels according to their importance. We divide the image into three kinds of region: high protection region, low protection region and non-protection region. UEP codes algorithm is used in our method to realize unequal protection levels and form the watermark bits for restoration. Because the high level bits have better distance characteristic in coding space, the high protection region can be restored more correctly than low protection region when the image suffers tamper and modification. In the current method, the protection for images only has two levels. Further improvements will include using better encoding and decoding algorithm to provide more degrees of protection and give better restoration capability for the images without introducing more redundant bits. ACKNOWLEDGMENT This work was supported by the Natural Science Foundation of China (60872116, 60832010, and 60773079), the High-Tech Research and Development Program of China (2007AA11Z247), the Shanghai Rising-Star Program (10QH14011), the Key Scientific Research Project of Shanghai Education Committee (10ZZ59), the Shanghai Specialized Research Foundation for Excellent Young Teacher in University (slg09005), and the OECE Innovation Foundation of USST (GDCX-Y-103).

432


Figure 4. Comparison results. The first column shows original images. The 2nd column shows the mask representing the protection levels and the 3rd column is the tampered images with noises. The 4th column presents the restoration results of the method using EEP code, while the 5th column is the results of our UEP based method. [7]

REFERENCES [1]

[2]

[3]

[4]

[5]

[6]

F. A. P. Petitcolas, R. J. Anderson, M. G. Kuhn, Information Hiding — A Survey, Proceedings of the IEEE, 87(7), 1999: 1062–1078. C. D. Vleeschouwer, J. F. Delaigle, B Macq, Invisibility and application functionalities in perceptual watermarking: An overview, Proceedings of the IEEE, 90(1), 2002: 64-77. P. W. Wong, N. Memon, Secret and Public Key Image Watermarking Schemes for Image Authentication and Ownership Verification, IEEE Transaction on Image Processing, 10(10), 2001: 1593-1601. S. Suthaharan, Fragile Image Watermarking Using a Gradient Image for Improved Localization and Security, Pattern Recognition Letters, 25(16), 2004: 1893-1903. H. Yang, A. C. Kot, Binary Image Authentication With Tampering Localization by Embedding Cryptographic Signature and Block Identifier, IEEE Signal Processing Letters, 13(12), 2006: 741-744. J. Fridrich, M. Goljan, Images With Self-correcting Capabilities, Proceedings of IEEE International Conference on Image Processing, 1999: 792–796.


[8]

[9]

[10]

[11]

[12]

[13]

X. Zhang, S. Wang, Statistical Fragile Watermarking Capable of Locating Individual Tampered Pixels, IEEE Signal Processing Letters, vol.14, 2007: 727–730. X. Zhang, S. Wang, Fragile Watermarking With Error-Free Restoration Capability, IEEE Transactions on Multimedia, 10(8) , 2008: 1490–1499. W. Feng, Z. Liu, Region-Level Image Authentication Using Bayesian Structural Content Abstraction, IEEE Transactions on Image Processing, 17(12), 2008: 2413–2424. S. Borade, B. Nakiboglu, L. Zheng, Unequal Error Protection: An Information-Theoretic Perspective, IEEE Transactions on Information Theory, 55(12), 2009: 5511-5539. C. Wang, M. Chiu, C. Chao, On Unequal Error Protection of Convolutional Codes From an Algebraic Perspective, IEEE Transactions on Information Theory, 56(1), 2009: 296-315. B. Masnick, J. Wolf, On Linear Unequal Error Protection Codes, IEEE Transactions on Information Theory, 3(4), 1967: 600-607. J. Wil, V. Gils, Two Topics on Linear Unequal Error Protection Codes: Bounds on Their Length and Cyclic Code Classes, IEEE Transactions on Information Theory, 29(6), 1983: 866-876.


Chuan Qin received the B.S. and M.E. degrees in electronic engineering from the Hefei University of Technology, Anhui, China, in 2002 and 2005 respectively, and the Ph.D. degree in signal and information processing from Shanghai University, Shanghai, China, in 2008. Since 2008, he has been with the faculty of the School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, where he is currently a Lecturer. His research interests include image processing and multimedia security. Qian Mao received the B.S. degree in Mechanical Engineering and Automation Science from Nanjing University of Aeronautics and Astronautics, Jiangsu, China, in 2000, and M.E. degrees in Traffic Information Engineering and Control from Shanghai Ship and Shipping Research Institute, Shanghai, China, in 2003, and the Ph.D. degree in Traffic Information Engineering and Control from Tongji University, Shanghai, China, in 2006. Since 2006, she has been with the faculty of the School of Optical-Electrical and Computer Engineering, University of Shanghai for Science and Technology, where she is currently a Lecturer. Her research interests include information theory and coding. Xinpeng Zhang received the B.S. degree in computation mathematics from Jilin University, Jilin, China, in 1995, and the M.E. and Ph.D. degrees in communication and information system from Shanghai University, Shanghai, China, in 2001 and 2004, respectively. Since 2004, he has been with the faculty of the School of Communication and Information Engineering, Shanghai University, where he is currently a Professor. His research interests include information hiding, image processing, and digital forensics.


433