Zero-Based Code Modulation Technique for Digital Video Fingerprinting In Koo Kang 1, Hae-Yeoun Lee 1, Won-Young Yoo 2, and Heung-Kyu Lee 1 1 Department of EECS, Korea Advanced Institute of Science and Technology, Guseong-dong, Yuseong-Gu, Deajeon, Korea (Republic of) ikkang@mmc.kaist.ac.kr 2 Digital Contents Research Division, Electronics and Telecommunications Research Institute, 161 Gajeong-dong, Yuseong-gu, Daejeon, 305-700, Korea zero2@etri.re.kr Abstract. Digital fingerprinting is a technique to protect digital contents from illegal reproduction and redistribution by marking unique information for individual user. A powerful but simple attack to diminish fingerprint signals is averaging. While several fingerprinting schemes against collusion attacks were proposed, they often do not fully account for multimedia data. In this paper, we propose a zero-based code modulation method which fully reflects marking assumption concept to embedding and detection of signals and anti-collusion code working mechanism as well. By manipulating 0-bit information of binary codes, detection accuracy and code separation under averaging attacks were enhanced. To demonstrate our method, we used an averaging-resilient fingerprint code based on GD-PBIBD theory and applied it video fingerprinting systems. Through the experimental results, we convince our method improves averaging resiliency of fingerprinting systems. 1 Introduction Digital fingerprinting is a copyright protection technique that traces illegal distributions of copyrighted contents. A unique fingerprint code that identifies a recipient is inserted into host contents. When copies of the content are found from illegal routes, the content seller can identify traitors who had distributed the content by extracting the embedded fingerprint code. This technique can be employed as a part system in DRM applications [1]. Averaging attack is a serious problem in digital fingerprinting applications. The averaging attack is an attempt to remove the inserted fingerprints by averaging several copies of the content. There are two main categories for averaging attack-resistant fingerprinting techniques. One is orthogonal modulation fingerprinting techniques; each user is assigned a spread spectrum sequence which is mutually orthogonal to each other as a fingerprint. The other approach is known as coded modulation methods that use collusionresistant fingerprint codes. Boneh et al. introduced a code working concept named Frame-proof code that prevents false alarm of recipients who did not join in averaging attacks up to c colluders [2]. Trappe presented a fingerprinting architecture that adopted the frame-proof idea using BIBD (Balanced Incomplete Block Design) R. Khosla et al. (Eds.): KES 2005, LNAI 3683, pp. 1108 1114, 2005. Springer-Verlag Berlin Heidelberg 2005
Zero-Based Code Modulation Technique for Digital Video Fingerprinting 1109 theory [3]. Most of coded modulation methods are working based on marking assumption which assuming colluders can only change fingerprint code bits in which they have different values [2]. Since most of existing coded modulation works focus on generic data, they do not explore special properties of multimedia data for fingerprint system design. In this paper, we propose a zero-based code modulation method against the averaging attack, which fully reflects the marking assumption to multimedia data. Our method modulates only 0 bit information in fingerprint code so that signal interruptions and signal strength diminution were eliminated. Our experimental results show the proposed method can efficiently embed and detect the fingerprints under averaging attacks along with multimedia data. This paper organized as follows. Section 2 addresses major features of orthogonal modulation methods and coded modulation methods to figure out their problems. In section 3, we describe our fingerprint code scheme and propose the zero-based code modulation method. Its effectiveness is demonstrated in Section 4 through experiments and conclusion is presented in Section 5. 2 Previous Fingerprinting Schemes 2.1 Orthogonal Modulation Method Orthogonal modulation approach is a straightforward method for digital fingerprinting. In this method, N orthogonal signals are used to accommodate N users, i.e. a mutually orthogonal signal is assigned to each user as a fingerprint. To decide the content owner or to trace colluders who made illegal copies, the same number of correlation as the number of users is required so that the detection complexity is high. An additional drawback is that when the number of contents involved in the averaging attack increases, the strength of orthogonal signals is attenuated deservedly so that the detector cannot correctly trace colluders. Assume that correlation is used as a detection statistics and M users join to the coalition to make illegal copy V. The signal of each user is orthogonal so that correlation values will be decreased in inversely proportional to the number of colluders M as follows. 1 1 L. C. = W[] i V[], i wherev = Cj. N M M (1) 1 = Wi [] Ci [] j NM This shows a limitation of detector when orthogonal signals are averaged and we can expect when the large number of users joins to make illegal copies, orthogonal modulation scheme will fail to retrieve colluders successfully. More detail study is well-defined in [4]. 2.2 Coded Modulation Method The coded modulation method was designed to accommodate more users than that of orthogonal method with the same amount of signals. Trappe proposed a representative coded modulation method [3]. The binary fingerprint codes derived from BIBD
1110 In Koo Kang et al. theory serve up to c-user averaging resistance. Averaging of up to c codes results in a unique code which can identify all codes associated with the averaging. A fingerprint signal for one user is composed of code bits and orthogonal signals as follows. w j = M i= 1 where b ij are {±1}, which is multiplied by u i according to i-th bit of j-th user code and u i are orthogonal signals for i-th bit position. At the detector, b ij are determined by correlation between w j and u i with a threshold value. When several fingerprint codes are averaged, the detected code is a bit-wise logical AND of those fingerprint codes. By comparing bit positions where the value is 1 in averaged code with a code book, the detector can find out which codes are involved in the alliance. One problem of this method is a difficulty to decide a threshold value. What the detector would concern is magnitudes of detected values. Let suppose 7 users join to the averaging; one code begins with (0,1, ) and all the others begin with (1,1, ), respectively. In order to trace colluders correctly, the colluded code should be recognized as (0,1, ) for the first two bits. However, extracted results could be different from the theoretical result in blind detection. As shown in Fig. 3, some correlation results for the first bit represent bigger values than that of second bit, and that means detector cannot decide the constant threshold value to separate two groups. Mess of correlation points results in incorrect binary result codes and detector will deservedly report innocent users as pirates. For this reason, the embedding method of equation 2 may not guarantee the marking assumption; the main working mechanism of averaging-resilient fingerprint code. b ij u i (2) 3 Zero-Based Code Modulation Technique 3.1 Anti-collusion Fingerprint Code and Embedding Anti-collusion fingerprint codes have a resiliency property for averaging attacks. Averaging attack on anti-collusion codes assumes that the result code is bitwise AND operation of those codes [3]; if at least one 0 is included in the codes at same bit positions, the result code bit at that position should be 0 no matter how many 1s are included in other codes. We used ( v, b, r, k, λ 1, λ2) GD-PBIBD (Group Divisible Partially Balanced Incomplete Block Design) theory to generate our test code for averaging attack-resiliency [5][6]. In ( v, b, r, k, λ 1, λ2 ) GD-PBIBD codes, the k indicates the number of 0s in a v-length code and also means the number of patterns to be embedded in our scheme. The k should be a smaller number than v [6] and that means our code modulation method is more efficient than that of conventional coded modulation method. For tests, we adopted (72,81,9,8,0,1) GD-PBIBD. The length of each fingerprint code is 72 bits and total 81 users can have a unique fingerprint code; each fingerprint code has bit 0s at different positions each other. This code scheme can trace up to 7 users when they deliver averaging attacks with their own codes. In a proposed scheme, the embedder modulates only 0 bit information in codes to fingerprint signals and inserts them to a host data. By utilizing only 0 bit information
Zero-Based Code Modulation Technique for Digital Video Fingerprinting 1111 in codes, the embedder handles only core factors of colluder tracing mechanism and the marking assumption. The way of M bits-length fingerprint code modulation is depicted in Fig. 1 and we embedded those signals to host data using commonly used spread spectrum embedding method [7]. Fig. 1. Zero-Based Code Modulation 3.2 Detection Process The detection steps are composed of three stages; estimation of fingerprint signals, examination of fingerprint codes and tracing of pirates. At the estimation step, the detector extracts expected noise-like fingerprint signals from corrupted contents. Because noise-like orthogonal signals are added to original contents, some noise-pass filters are available. At the next step, our detector examines which indexes patterns are embedded in the estimated pattern through a normalized correlation strategy [7] as depicted in Algorithm 1. Since our fingerprint code has fixed number of 0 bit information, k, what the detector has to know is indexes of embedded signals. In case of the number of 0s are over the number k, the content must have been attacked by the way of averaging because the codes cannot have k number of 0 bits at exactly same positions. At the final step, we can find out the pirates by examining the fingerprint code book and the result code. The colluder tracing method is the same way as that of conventional system [3]. 1. Initialize result code bits R 1, R 2,..., R M to 1 2. For estimated signals E 1, E 2,...E K and detector generated signals S 1, S 2,..., S M - calculate correlation between E i and S j (1 I K, 1 j M) - if correlation > threshold then set R j to 0 and resume correlation between E i and S j+1 - else if correlation < threshold then resume correlation between E i and S j+1 Algorithm 1. Fingerprint detection and code decoding 4 Experimental Performance 4.1 Code Separation Experiments We will show experiments of detection value separations and results. This experiment is aimed to see how well the detection values are separated because the reason marking assumption does not work on multimedia is due to unclear separation of detected values. For this purpose, we made a fake code to demonstrate the worst cases of averaging attacks as shown in Fig 2. In a case of 0 bit detection tests, only one, among n
1112 In Koo Kang et al. bits at the same bit position in n codes, has a bit 0 and the others have 1. Theoretically, all bits of averaged code are 0. We illustrated detection results of both systems in Fig 3. In (a), (c) and (e), we can see the detected values are getting blended up as the number of collusion increases. This situation happens because as the signals for bit 1are increased in equation 2, the linear correlation values are also increased. However in the proposed scheme, correlation points are clearly divided into two groups and the threshold can be easily determined. Even though detected values for bit 0 are decreased as collusions are increased, we can see two obvious partitions even in 7 collusion case. Therefore, our proposed scheme enhances code detection capabilities to decode the averaged code and to capture the colluders who delivered averaging attacks on multimedia data. Test code for bit 1 detection for a 7 averaging attack (7 codes of 72 bits) user 1~n : 1,1,1,1,1,1,...,1,1,1,1,1,1 Test codes for bit 0 detection for a 7 averaging attack (7 codes of 72 bits) user 1 : 0,1,1,1,1,1,1,0,1,1,1,1,1,1,0,1,1,...,0,1,1,1,1,1,1 user 2 : 1,0,1,1,1,1,1,1,0,1,1,1,1,1,1,0,1,...,1,0,1,1,1,1,1 user 3 : 1,1,0,1,1,1,1,1,1,0,1,1,1,1,1,1,0,...,1,1,0,1,1,1,1... user n : 1,1,1,1,1,1,0,1,1,1,1,1,1,0,1,1,1,...,1,1,1,1,1,1,0 ---------------------------------------------------------------------------------------------------- averaged : 0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,0,...,0,0,0,0,0,0,0 Fig. 2. Fake codes for 0 and 1 bit detection experiments 4.2 Colluder Tracing Experiments on Video We applied our approach to video data extracted from DVDs and measured the performance of tracing colluders. We used fingerprint code from (72,81,9,8,0,1) GD- PBIBD and CIF (352x288) sized video data as host signals. For the video fingerprinting performance assess, we averaged uniquely fingerprinted contents from 2 collusions to 7 collusion cases 100 times, respectively, and extracted the colluders. Table 1 shows the result of test; the number of correctly extracted colluders for each collusion case. When the number of colluders was below 4, we could trace all colluders successfully. When the colluder number exceeded 4, our detector partially failed to catch all colluders, but could trace at least one user and there was no report on innocent users. Table 1. Detection results from 10 illegal copies of 10 videos for each collusion case Number of extracted colluders 1 2 3 4 5 6 7 1 collusion 100 - - - - - - 2 collusion - 100 - - - - - 3 collusion - - 100 - - - - 4 collusion - - - 100 - - - 5 collusion - - - 3 97 - - 6 collusion - - - 1 3 94-7 collusion - - 1 1 2 6 90
Zero-Based Code Modulation Technique for Digital Video Fingerprinting 1113 Fig. 3. Correlation value separation results both systems: detected correlation values for 0 bit (black colored shape) and 1 bit (white colored shape). (a), (c) and (e) are result of legacy systems and (b), (d) and (f) are of the proposed system 5 Conclusions The anti-collusion fingerprint code scheme has a property of the colluder tracing function based on a marking assumption concept. The marking assumption, however, does not work well in multimedia data because the detection scheme of fingerprint signals works with correlation-based method and the results are not always agree with its theoretical result. This disagreement cause wrong detections of innocent users,
1114 In Koo Kang et al. which is the most serious problem in fingerprinting applications. To get rid of this trouble, we proposed a zero-based code modulation method which fully reflects running mechanisms of marking assumption and anti-collusion codes. We modulated and embedded only 0 bit information of binary codes so that correlation results for bit 0 and 1 could be separated as maximally as possible. Through code bit separation experiments, we convinced our proposed method is effective to classify two groups of correlation points even though the worst cases of averaging attacks. From video data experiments, we confirmed that the bit separation property was maintained and the detector could trace colluders successfully in the corrupted contents. References 1. Junseok Lee, Seong O.H.: A DRM Framework for Distributing Digital Contents through the Internet, ETRI Journal, 2003, pp. 423-436. 2. D. Boneh, J. Shaw: Collusion-secure fingerprinting for digital data, IEEE Trans. On Information Theory, Vol. 44. (1998) 1897-1905 3. W. Trappe, M. Wu, Zhen Wang, K.J.R. Liu: Anti-collusion Fingerprinting for Multimedia, IEEE Trans. On Signal Processing, Vol. 51. (2003) 1069-1087 4. Z. Jane Wang, Min Wu, Hong Zhao, K. J. Liu and Wade Trappe: Resistance of orthogonal Gaussian fingerprinting to collusion attacks, Proc. of ICASSP, pp 724-727, Apr. 2003 5. C. J. Colbourn, J. H. Dinitz: The CRC Handbook of Combinatorial Designs, CRC (1996) 6. Willard H. Clatworthy: Tables of two-associate-class partially balanced designs, National Bureau of Standards Washington D.C. U.S. (1973) 7. I. J. Cox, M.L. Miller, J.A. Broom: Digital Watermarking, Morgan Kaufmann Publishers: San Francisco CA (2002)