A SEMI-FRAGILE WATERMARKING SCHEME FOR IMAGE AUTHENTICATION AND TAMPER-PROOFING

Transcription

1 A SEMI-FRAGILE WATERMARKING SCHEME FOR IMAGE AUTHENTICATION AND TAMPER-PROOFING Nozomi Ishihara and Kôi Abe Department o Computer Science, The University o Electro-Communications, Chougaoa Chou-shi, Toyo, Japan Phone: Fax: {n-ishi, abe}@cacao.cs.uec.ac.jp Abstract This paper describes a semi-ragile watermaring scheme or image authentication and tamper-prooing. Each watermar bit is duplicated and randomly embedded in the original image in the discrete wavelet domain by modying the corresponding image coeicients through quantization. The modication is made so that it has little eect on the image and the watermaring is robust against tampering. The watermar bits or authentication are obtained by a majority vote on the extracted bits. The bits that lose the vote are treated as having been tampered with, and the location o the bits indicates the positions that have been tampered with. Thus, authentication and tamper-prooing can be done by observing images o watermars that win and lose votes. The proposed scheme is robust against JPEG compression or acceptable modication, but sensitive to malicious attacs such as cutting and pasting. Key words: watermaring, DWT, semi-ragile, authentication, tamper-prooing 1. Introduction The success o the Internet has brought about substantial beneits. Multimedia distribution is one o them. But copyright protection or tamper-prooing has introduced new challenges because multimedia data is easy to copy or mody. Watermaring is one solution to this problem. Several digital watermaring algorithms have been proposed. Among them is a robust watermaring system that embeds a solid watermar and protects copyright [1], []. But it is hard or robust watermaring to respond to a wide variety o possible distortion including malicious modications. To solve this problem, ragile watermaring that can detect what ind o distortions happened to the image was proposed. In this watermaring algorithm the watermar is wealy embedded so as to be sensitive to distortion. The receiver can chec the integrity o the received image by extracting and verying the watermar. A method based on this algorithm was proposed by Kunder and Hatzinaos [3] who used a tamper-assessment unction (TAF) to chec the integrity o images. The method is eective or detecting various inds o possible distortion. However, because o the weaness o the watermar it may collapse even under an allowable modication such as JPEG compression. A semi-ragile watermaring algorithm was then proposed to overcome this problem. Semi-ragile watermars are tolerant to allowable modications but sensitive to malicious attacs. Zhou [4] used error-correction coding (ECC) to develop a semi-ragile watermar. It was, however, reported that the algorithm might be susceptible to attacs that changed the maximum or minimum values in 8*8 blocs. Ko [5] and Chi [6] also developed semi-ragile watermaring schemes, but these schemes use the discrete cosine transorm and are thereore not compatible with the JPEG 000 standard [7]. These schemes also determine the coeicients ater decomposed into requency domain, which yields high overhead. In this paper, we propose a semi-ragile watermaring algorithm with low overhead that detects the position o distortions and shows it via an image. This algorithm uses Discrete Wavelet Transorm (DWT) and duplicates the watermar image and embeds the resulting copies to the original image according to the pre-determined ey. It can chec the authority and the tampering by image without the availability o the original watermar by taing vote. In addition, we use an embedded method that is robust to attacs and has less impact on the original image.. Wavelet Transorm o Images To transorm image to requency band, we use DWT to mae it compatible with JPEG 000 standard. To perorm the orward DWT, the JPEG 000 standard uses a one-dimensional (1-D) sub-band decomposition o a 1-D set o samples into low-pass and

2 high-pass samples. Low-pass samples represent a down-sampled low-resolution version o the original set. High-pass samples represent a down-sampled residual version o the original set. When the two-dimensional DWT (DDWT) is done, the image is decomposed into hierarchical sub-bands. The basic idea o the DDWT is to decompose the requency into our parts (i.e., LL1, HL1, LH1, and HH1 sub-bands) by sub-sampling the horizontal and vertical channels using sub-band ilters [8]. The sub-bands labeled HL1, LH1, and HH1 represent wavelet coeicients at the inest scale. To obtain wavelet coeicients on the next coarsest scale, the sub-band LL1 is urther decomposed and sub-sampled. This process is repeated several times, as determined by the specic application. In JPEG000 standard Le Gall reversible 5-tap / 3-tap and Daubechies irreversible 9-tap/7-tap ilters are supported or transormation. In this paper, we used the ormer ilter to ensure less impact on the watermar. Ater the watermar is embedded, these DWT coeicients are reconstructed by an inverse DWT (IDWT). 3. Proposed Method 3.1 General description Our proposed method duplicates a binary watermar image w which is to be embedded in the original 3-level DWT domain image. To embed the duplicated bits, we use the ey cey, which indicates the embedded leve sub-band, and position. These parameters are pre-deined randomly. We also use the quantization parameter Δ, and the number o duplicated bits cnum. We also use LL sub-band to countereit the collage attac [9]. To countereit the attac, the embedded bit has to be correlated with other bits. So we correlates the embedded bit to the corresponding LL band s positions in DWT level 3. In Section 3. we describe the detail. To authenticate the watermared image, we irst extract the embedded watermar bits by reerring cey, Δ, and cnum, just as the reverse o the embed process. Then we tae a vote or reconstructing the watermar image w. By checing w, we authenticate who produced the image. Next, we plot the bits which lost the vote to the tamper-detection image. Tamper-detection image has the same size o the watermared image, and is decomposed by 3-level DWT. By plotting the bits to the originally embedded position, we chec the tampered place. Ater plotting, the tamper-detection image is reconstructed by IDWT. The reason why we reconstruct the tamper-detection image is to spread out the tampered watermar to mae the tampered-place more visible. Note that we don t need the original watermar image in extraction process by using voting system. In our scheme which chooses the position randomly, some embedded watermar bits may be ragile compared with other semi-ragile schemes which choose the robust position careully. However, by duplicating watermar image bits and taing a vote among the copies or reconstructing the watermar image, the watermar bits which are embedded into robust position are liely to win ragile bits. As a result, a suicient number o watermar bits are embedded, this scheme is robust against non-malicious attacs. We also note that our scheme has low overhead because we can pre-determine parameters or the embedding process. 3. Correlation with LL band or countereiting the collage attac In Section 3.1, we briely described the method against the collage attac. In this section, we describe the detail. To correlate the embedded bit with other bits, we XOR the average o adjacent our pixels in level 3 LL band with the bit to be embedded. To choose the adjacent our pixels, we mae use o wavelet parent-child relationship [1]. The steps are described as ollows: (See Fig. 1.) 1) For each embedded position (x, y), we choose the corresponding po- x, in level 3: sition ( ) LL y LL ( x LL, y LL ( x 4, y 4), ) = ( x, y ), ( x, y), (x,y) is in level1 (x,y) is in level (x,y) is in level3 ) The adjacent our pixels Fig. 1 Correlation with LL band. ( x LL, y LL), ( x, +1) LL y LL, ( x ) LL + 1, y LL, ( x LL + 1, yll + 1) are averaged and quantized. The value is then XORed with the embedded bit as described in the ollowing subsection.

3 3.3 Embedding process The embedding process consists o the ollowing ive steps: (See Figs. and 3.) 1) The watermar image w (binary image, n msize), number o duplicated bits cnum, and quantization parameter Δ are determined by the user. I the amount o duplication is increased, the watermar image is more robust to distortion, enabling minute distortions to be detected. I Δ is increased, the watermar bit is embedded more solidly. When the value o cnum is ixed, the elements o the array cey with the size o Nw = wx wy cnum are randomly generated. Each element o cey indicates embedding leve sub-band (except LL3), and position. ) The original image is decomposed by 3-level DDWT (-dimensional DWT). 3) Reerring to the cey, the position (x, y) and the quantized watermar bit are determined. We describe the quantization process in Section 3.4 4) The average o adjacent our pixels in LL is quantized and XORed with the embedded bit. The result is embedded to the position determined in step 3. We describe the details o embedding method in Section ) Ater repeating the steps 3 and 4 by the length o cey, the image is reconstructed using IDWT. Fig. Duplication o watermar. Fig. 3 Embedding process. 3.4 Quantization process and embedding method Let a wavelet-transormed image be denoted by a set o integer unctions l,, where l, { lh, h hh}, and ( m, denote a leve sub-band, and position, respectively. To embed a watermar, the quantization unction Q ( ) which maps an integer to a binary value is used, which is deined as ollows: Q( ) = 0, 1, r r < ( r+ 1), < ( r+ 1), where Δ is the quantization parameter. Fig. 4 illustrates the mapping. r = 0, ±, L r =± 1, ± 3, L Let the i-th watermar bit be denoted by w(, i= 0, L, Nw 1. In Kunder and Hatzinaos' method [3], Δ is simply added to the wavelet coeicients during the embedding process. However, this process has a large eect on the image and maes it susceptible to distortion. In this paper, we improve this method by maing the embedded value to tae the median o the interval. (See Fig. 5) 1) I Q( l, ) = w(, then mody l, so that the value be the median o that interval: A) I l, 0, l : = + mod F,, B) Otherwise, l, : = mod F,, where mod F denotes mod. l, ) Otherwise, mody l, so that Q( l, ) = w( Fig. 4 Quantization unction Q ( ). and the value o l, be the median o the interval:

4 A) I l, 0 B) Otherwise, : = 3 + mod F, mod F+, mod F > mod F : = 3 mod F, + mod F+, mod F > mod F 3.5 Extracting watermar bit and detecting the tampering The process or extracting the watermar bits and reconstructing the watermar image is as ollows: 1) The possibly tampered image ' is decomposed by 3-level DWT to obtain. ' ) Reerring to cey, Q( ' l, ) is calculated and XORed with the quantized average o adjacent our pixels in LL band just lie the embedding process. Thus the watermar bits w '( are extracted. 3) From the extracted watermar we gather the cnum pixels which were duplicated rom the same position in the embedding process, and tae a vote on the pixels or the watermar image. The reconstructed image w which indicates the authorization is then checed. The process o detecting distortion is as ollows: Fig. 5 Quantization process (in case o > 0 ) 1) Reerring to the result o the above step 3, which the watermar bits w ''( lost the vote is determined. ) These lost bits are plotted on a blan image in 3-level DWT domain at the w ''( position. 3) The resulting image is reconstructed through IDWT. The extracted watermar image w' and tamper-detection image w' ' produced are checed to see the image is authentic. These processes are illustrated in Fig Experimental results 4.1 Characteristics o our scheme Fig. 6 Extraction and tamper-detection process. We used a sized gray-scale image as the original image. Because the watermar image is intended to indicate the ID o the author or receiver clearly, we used the name o one o us as the 3 3 sized binary watermar image. Let the duplication o the w x wy size watermar image be "total watermar bits". First, we examined the relations o PSNR and cnum, and PSNR and Δ, where PSNR is deined to be 10 log RMSE, RMSE being the root-mean-square error between a watermared (or an attaced watermared) image and the original image Bridge. Next we modied the watermared images Bridge, Lena, and Boat by sotness so that the modied image s PSNR be 39.96, 43.63, and 4.9, respectively. We then examined the corruption probability o the total watermar bits and the watermar image. The results are shown in Fig. 7 and 8, and Tables 1 and. Figures 7 and 8 indicate that the more Δ or cnum is increased, the more PSNR is decreased. The tables show that the greater the increase in cnum, the more robust is the watermar image. This is because increasing the number o non-broen bits leads to accurate voting results. But in case o Δ = 5 (the embedding is too wea), the watermar image was severely broen even cnum is increased. I Δ is increased, the total watermar bits become more robust. These results indicate that there is a tradeo between Δ, cnum, and the PSNR: the more Δ or cnum is increased, the more robust or sensitive to distortion the watermar is, but the more severely the original image is distorted. Our algorithm taes a vote to detect distortions. I the distortion is mild, the vote may be accurate, but the distortion is severe, the number o broen watermar bits will increase and it will become more dicult to detect distortions accurately. We thereore examined the relation between detection accuracy and the ratio o broen watermar bits to the total bits when the noise was enhanced gradually. The results are shown in Fig. 9, where the total watermar bits are 0,480, and the broen rate was varied by introducing Salt and pepper noise. The result shows that our scheme detects distortion accurately even when the rate was nearly 0%. The accu-

5 racy remains 95% up to the broen rate o 40%. PS NR cnum Fig. 7 cnum vs. PSNR PS NR Δ Fig. 8 Δ vs. PSNR Table 1 Ratio o bits broen by sotness to total watermar bits (%). Δ Cnum Bridge Lena Boat Bridge Lena Boat Bridge Lena Boat Bridge Lena Boat Cnum 4. Sensitiveness and robustness evaluation Table Ratio o area broen by sotness to the total area o watermar image (%). In [10], the attacs were categorized into malicious and non-malicious. For example, mild iltering such as JPEG compression is non-malicious, while geometric transorm such as cut & paste, translation or rotation is malicious. So we employed JPEG compression, Median ilter, and salt and pepper noise as benchmars which examine the robustness against non-malicious attacs, and cutting as the malicious attacs to test the sensitiveness. We show irst the sensitiveness against the malicious attacs. Next, we evaluate our scheme with respect to the robustness against the non-malicious attacs by comparing with other schemes. Fig. 9 Accuracy o detection Sensitiveness against malicious attacs We examined what happened when the watermar image was cut ater being compressed under a JPEG compression rate o 30% to show the eectiveness o our scheme. The results are shown in Figs. 10, 11, and 1. In the tamper-detection image, the white and blac areas are the non-broen and broen areas, respectively. These results indicate that a watermared image is compressed under a high-quality JPEG and part o the image is then altered, our scheme can detect the distortion accurately. Our scheme was also ound to be able to accurately detect the distortion even the order o the modication and tampering operations is reversed. Δ Bridge Lena Boat Bridge Lena Boat Bridge Lena Boat Bridge Lena Boat Comparison with other schemes about robustness against non-malicious attacs We compared our scheme with one o the study [3] previously proposed or detecting tampering via an image. The comparison was made by examining the rate o broen bits to the total number o watermar bits under the non-malicious benchmars described

6 above. The results, which are shown in Table 3, are the average or 10 watermared images. Note that the scheme [3] uses a Haar wavelet or reversible transormation, while we used a Le Gall 5-tap/3-tap ilter. In the tables, 30% and 50% indicate the JPEG compression rates, Noise indicates 1% salt and pepper noise, and Med indicates the median ilter. The results show that our proposed method is robust against allowable modications. This is because our method embeds a watermar by modying coeicients so as to be the median o the interva while in the method [3] a watermar is simply embedded by adding Δ. (a) (b) Fig. 10 Lena: (a) original image; Fig. 11 Watermar image Fig.1 Tamper- (b) tampered image ater watermared. (a) original; (b) extracted. detection image. 5. Discussion Table 3 Rate o the number o broen Our proposed method detects distortion via randomly embedded water- watermar bits to the total number o watermar mars. So an insuicient number o watermars are embedded, we can- bits under the non-malicious benchmars. not obtain an accurate result. In this study, on the basis o the tradeos indicated Attac in Figs. 7 and 8, and Tables 1 and, and the density o the broen watermar bits in Fig. 1, we see that when the number o embedded watermar bits is at least 0,000 with o at least 5, the method is robust enough against non-malicious modication and sensitive to distortion. Re.[3] Proposed method 30% % Noise Med We also note that real-time processing is required or such applications as medical or accident images used or the court evidence. Hardware implementation o semi-ragile watermaring is attractive or applications using digital cameras. Our scheme would be well suited or the purpose. 6. Conclusions In this paper, we proposed a semi-ragile watermar scheme that detects distortion and provides a visual method o authentication in low overhead. The experimental results showed that (1) the watermar bits were embedded solidly, () tampers could be detected visually, (3) our scheme was robust against allowable modications but sensitive to malicious attacs, (4) our scheme has low overheads because the embedding parameters can be pre-determined. 7. Reerences [1] M. S. Hsieh, D. C. Tseng, "Hiding Digital Watermars Using Multi-resolution Wavelet Transorm," IEEE Trans. Ind. Electron, Vol. 48, No. 7, pp , Oct [] Y. H. Seo, S. Y. Choi, S. H. Par, D. W. Kim, "A Digital Watermaring Algorithm Using Correlation o the Tree Structure o DWT Coeicients," IEICE Trans. Fundamentals, Vol. E87-A, No. 6, pp June [3] D. Kunder, D. Hatzinaos, "Digital Watermaring or Telltale Tamper Prooing and Authentication," Proc. IEEE Vol. 87, No. 7, pp , July [4] X. Zhou, X. Duan, D. Wang, "A Semi-Fragile Watermar Scheme For Image Authentication," Proc. IEEE 10th Int. MMM'04, pp , 004. [5] C. C. Ko, C. H. Huang, "A Novel Semi-Fragile Watermaring Technique For Image Authentication," Proc. 6th IASTED International Conerence on Signal and Image Processing (SIP 004), Honolulu, Hawaii, pp. 4-9, Aug [6] K. H. Chi, L. C-Tsun, Semi-Fragile Watermaring Scheme or Authentication o JPEG Images, Proc. o ITCC 04, Vol. 1. pp. 7-11, April [7] A. Sodras, C. Christopoulos, and T. Ebrahimi, The JPEG000 Still Image Compression Standard, IEEE Signal Processing Magazine. Vol. 18, No. 5, pp36-58, Sept [8] R. K. Young, Wavelet Theory and its Applications, Kluwer Boston, MA, USA, 1993 [9] M. Holiman and N. Memon, Countereiting Attacs on Oblivious Bloc Wise Independent Invisible Watermaring Schemes, IEEE Trans. Image Processing, Vol. 9, No.3, pp , 000. [10] E. Ozgur, S. Bulent, C. Baris, N. Umut, A. Mahmut, Comparative evaluation o semragile watermaring algorithms, J o Electronic Imaging Vol. 13, No. 1, pp , Jan. 004.