EXPOSING THE DOUBLE COMPRESSION IN MP3 AUDIO BY FREQUENCY VIBRATION. Tianzhuo Wang, Xiangwei Kong, Yanqing Guo, Bo Wang

Transcription

1 EXPOSIG THE DOUBLE COMPRESSIO I MP3 AUDIO BY FREQUECY VIBRATIO Tianzhuo Wang, Xiangwei Kong, Yanqing Guo, Bo Wang School of Information and Communication Engineering Dalian University of Technology, Dalian, 64, China ABSTRACT A novel approach is presented to detect doule compressed MP3 audio y frequency viration. With the analysis of doule compression effect on MDCT (Modified Discrete Cosine Transform coefficients in MP3 audio, we propose a simple feature called FVV (frequency viration value to measure the viration caused y doule compression. The experimental results on challenging dataset show that our method outperforms most of the existing methods in doule MP3 compression detection, especially with a second itrate higher than the first one. Besides, we can also estimate the original it-rate for a doule compressed MP3 y this technique. Index Terms MP3, audio forensics, doule compression detection, frequency viration, MDCT coefficients. ITRODUCTIO The past decades witness the fast development of digital and networ technologies, which had een integrated into people s life, and ecome an indispensale part. People are ale to record the information in everyday life with the increasing popularity of moile terminals. And everyone could share the convenience gifted y digital multimedia. However, the authenticity of digital multimedia is difficult to e guaranteed with the occurrence of intelligent editing software, and people can easily tamper the digital multimedia without professional training. As a part of multimedia, MP3 audio is one of the most commonly used audio format, and some methods of authenticity detection [][][3][4] have een proposed in recent years. However, it also faces some challenges and some prolems cannot e solved perfectly. Doule compression is used frequently in the process of tampering in MP3 audios, as the audio must e saved after modifying. Doule compression detection is then put forward, which can assign two inds of issues [5] : Up-transcoding (corresponding to the second compression using a higher quality than the first one and down-transcode (corresponding to the second compression using a lower quality than the first one. Due to the complexity of the MP3 coding process, it is very difficult to detect doule compression of audios in MP3 formats. Yang et al. [6] used the numer of smallvalue MDCT coefficients as features to detect the audio authenticity. It is ased on the oservation that the coefficient distriution is variant etween the fae-quality MP3 and a normal one. Still, in [7], Yang et al. extracted the first 9 ins of quantized MDCT coefficients as features, and put these features into SVM classifier. In particular, gloal distriution, as well as and distriution is also exploited to distinguish the normal MP3 from a doule compressed one. Though this method has an impressive result in detecting an up-transcoded audio, it has a limited performance in detecting a down-transcoded one. M. Qiao et al. extracted statistical features of quantized MDCT coefficients of MP3 audio, and exploited these features in a discriminative model [8][9][5]. However, these methods ased statistical features are limited in application as their high computation complexity. Bianchi et al. [] proposed a method of doule compressed detection considering simulated normal MP3 file, which is achieved y removing a given numer of PCM samples of decompressed audio file and recompressing the remaining samples to the same compression quality of the file. However this method still needs other statistical features to wor together, which is time-consuming. In this paper, we present a novel approach with sincerely respect to detecting doule compression for oth up-transcoded and down-transcoded MP3 audios, and estimate the original it-rate for the first compression. As we now, the ey issue for doule MP3 compression detection is to develop effective features which are discriminative etween oth inds of audios. Previous methods focus on the properties of MDCT coefficients in spatial domain, ut we find that these coefficients in transform domain have etter and more regular distriution. We extract a feature vector called FVV (frequency viration value to measure the viration caused y doule compression, and give a more effective detection. Experimental results show that our method outperforms the state of-the-art method in oth up-transcoding and downtranscoding cases. The rest of the paper is organized as follows. The detailed description of the proposed algorithm is shown in Section ; experiments and discussion are presented in Section 3, followed y conclusion in Section /4/$3. 4 IEEE 45 ChinaSIP 4

2 . PROPOSED ALGORITHM. Analysis of doule MP3 compression As we now, doule MP3 compression contains doule quantization on the MDCT coefficients, which maes the MDCT coefficients different from that of the normal one. The coefficient of a normal MP3 approximately oeys a Laplace distriution []. After the doule compression process, this distriution can t e ensured, and this property mae it possile to distinguish a normal MP3 audio from a doule compressed one. It is not easy to tell whether a coefficient distriution oey a Laplace distriution of a normal map3 audio. Figure shows the coefficient distriution of a normal MP3 and a doule compressed one. The figure demonstrates that the coefficient distriutions of oth audios are nearly the same, though with some minor differences in local. (a. The normal MP3 audio (. Doule compressed MP3 Figure MDCT distriution in transform domain (a normal MP3 ( doule compressed MP3 Figure MDCT coefficient distriution In view of this case, we consider to descrie the coefficient distriution in a transform domain, for the reason that an orthogonal transform (e.g. DFT can re-distriuted the energy, and some good properties occur. We otain two inds of the quantized MDCT coefficients which are respectively a normal and a doule compressed one with Hoffman decoding. Then MDCT coefficient distriution is computed, which is then transformed in a DFT method. We show the coefficient distriution in transform domain in Figure to find out the answer whether they are still discriminatory if looing in transform domain. In Figure, the normal MP3 audio has a smooth distriution in side loe, contrary to the distriution in a doule compressed MP3 audio with apparent viration. So it is possile to detect doule compression utilizing the volatility of the MDCT coefficient in transform domain.. Calculation of frequency viration value As the quantized MDCT coefficients has discrete value, its discrete Laplace distriution can e rewritten as n xa ( exp( ( n is the quantization threshold etween - to. The causal sequence of x a (n can e denoted as: n exp( n x( x a ( n n (3 exp( n Then the DFT of it is as follows: X ( x( W x ( n W n ( n xa ( exp( ( x ( exp( a n a n ( exp( n exp( Here we denote g exp( /, and Eq. (4 can e rewritten as: X ( ( g n exp( (4 After the oservation and analysis, we certify the proposed method through mathematic derivation, and then calculate the feature vector called FVV (frequency viration value to measure the viration. In general, a typical continuous Laplace distriution can e defined y xa t ( t exp( ( g g ( g cos jg sin( g g cos( (5 45

3 It is clear that x ( has a smooth amplitude-frequency curve. While, in a doule compressed audio, the quantized MDCT coefficients distriution is different from Eq. (, i.e. coefficients in high frequency are lost. In other words, quantized MDCT coefficients distriution is cut off in the edge. To simplify analysis, we use the distriution multiplied y an orthogonal window to simulate this process. An orthogonal window can e rewritten as A, t h ( t (6, t Discrete sequence h (t and causal discrete sequence h (t are donated as follows: d A, t h ( (7 d, t h ( A d n d (8 So the distriution after cutting off is as follows x ( x( h( (9 X ( X ( * H ( ( Where * means convolution, H ( is the DFT of h (. H ( h ( W ( h( W h( n W sin( d / cos j sin sin( d / sin( / It is ovious that distriution has viration in side loe. As a result, convolution X ( of X ( and H ( will also has apparent viration in side loe. In conclusion, we can directly use the volatility of the MDCT coefficient in transform domain to judge a doule compressed MP3. We propose T to donate the frequency viration value (FVV, and it is defined y m T ydx y' dx Y ( i Y '( i ( i ( Where y is the MDCT coefficient in frequency domain, y is a fitting line of the curve, m is the numer of fitting, and T is the area formed y the y and y '. Y (i and Y ' ( i are the discrete form of y and y '..3 algorithm pipeline Recognizing the properties of MP3 doule compression, we are going to present our proposed method ased on the frequency viration value. The diagram of the method is shown in Figure 3. The steps of algorithm implementation are as follows: (a Decompress the MP3 audio file to get the quantized MDCT coefficients; ( Transform it y DFT; (c Calculate the FVV; (d Determine doule compression. ote that the parameter Th is only related to the property of normal and doule compressed MP3 files. Its value can e predetermined through experiment with any fixed first itrate and second itrate. Figure 3 The diagram of the detecting method 3. EXPERIMETAL RESULTS AD AALYSIS An audio dataset has een uilt to validate the ideas proposed in previous sections. It contains all 4 audio files, each of which in MSWAVE format and 5s long []. The audio files elong to four different categories: speech alone: containing dialogues; music alone: containing music only; 3 speech-over-music: containing speech with music; 4 environmental sound: the files relative to recording environmental. Each of WAV file has een compressed with it-rate BR chosen in [64, 96, 8, 6, 9] Kit/s. These single compressed MP3 files have een decompressed and compressed again with it-rate BR chosen in [64, 96, 8, 6, 9] Kits/s. 3. Effect of our proposed feature To prove the effect of frequency viration value, we computed the values assumed y the proposed feature, which is shown in Figure 4, where the horizontal axis represents sequence numer of MP3 audio files, and the vertical axis represents the feature proposed. It shows x normal 9ps 96ps-9ps ps-9ps 8ps-9ps 6ps-9ps 9ps-9ps 4ps-9ps 56ps-9ps 3ps-9ps 样本数量 Figure 4 Comparison of the proposed feature 45

4 comparison of the proposed feature etween normal 9Kps MP3 and 8 types of recompressed MP3. ormal 9Kps means normal MP3 with the it-rate 9Kps. 96Kps-9Kps means doule compressed MP3 transcoded from 96Kps to 9Kps and so on The feature of normal MP3 are all near to zero, whereas those of doule compressed MP3 are all lager than zero. By comparing the feature with a threshold, it is possile to discriminate these two inds of files: doule compressed files and normal ones. 3. Doule Compression Detection To prove the effect of proposed method ased on the viration frequency, we set up experiments using different methods. The results are shown in Tales,,3,4, where BR means the first itrate and BR means the second itrate. The Tale Detection accuracy of our method BR vs BR Tale Detection accuracy of [7] BR vs BR Tale 3 Detection accuracy of [5] BR vs BR Tale 4 Detection accuracy of [] BR vs BR The results in tale show that our method is accurate and stale to detect doule compressed MP3 files. In the comparison experimental, we introduce Yang et al. Method [7], Qiao et al. Method [5] and Bianchi et al. Method [], and the results are shown in Tale, 3, 4. Figure 5 shows detection accuracy of different method with BR=6Kps. With BR declining, all of the four methods get worse. But our method still achieves the etter performance. Besides, the proposed wor with only one-dimensional feature. ot only the method solves the prolem of down-transcode much etter, ut also has higher detection accuracy in up-transcode. Accuracy[%] Proposed Yang et al qiao et al Bianchi et al Different compression Figure 5 Detection accuracy of different method with BR=6Kps 3.3 Reveal the Real Bit-Rate We utilize a earest eighor classifier to reveal the real it-rate. Tale 5 shows the confusion matrix of average multi-classification results. R means the real first compression it-rate of doule compressed MP3. E means the estimation of the first it-rate. Tale 5. The estimate result E vs R The estimate result shows that the proposed method could reveal the real it-rate of doule compressed MP3. Similar to the results of doule compression detection, the performance of revealing the real it-rate is etter in the uptranscoded than in down-transcoded. 4. COCLUSIOS In this paper, we present a novel and simple method to detect doule MP3 compression ased on the frequency viration of MDCT coefficients. To estimate the effect of doule MP3 compression on the quantized MDCT coefficients, we ring the frequency viration value and prove its reasonaility. Experimental results show that our method is very effective to detect doule compression than the typical method in the field. Besides, we can also estimate the original it-rate for a doule compressed MP3. However, the detection of down-transcoding is still an issue for more people to loo into it. 453

5 5. REFERECES [] R. Yang, Z. Qu, and J. Huang. Detecting digital audio forgeries y checing frame offsets. In ACM Multimedia and Security Worshop, pages 6, 8. [] Maher R. Audio forensic examination [J]. Signal Processing Magazine, 6(: 84-94,9. [3] Koenig B E, Lacey D S. Forensic authentication of digital audio recordings [J]. Journal of the Audio Engineering Society, 57(9: ,9. [4] Rodríguez D P, Apolinário J A, Biscainho L W P. Audio authenticity: Detecting EF discontinuity with high precision phase analysis [J]. IEEE Transactions on Information Forensics and Security, 5(3: ,. [5] Qiao M, Sung A H, Liu Q. Improved detection of MP3 doule compression using content-independent features [C]. In ICSPCC, 3. [6] Yang R, Shi Y Q, Huang J. Defeating fae-quality MP3 [C]. Proceedings of the th ACM worshop on Multimedia and security. ACM, 7-4, 9. [7] Yang R, Shi Y Q, Huang J. Detecting doule compression of audio signal [C]. International Society for Optics and Photonics,. [8] Liu Q, Sung A H, Qiao M. Detection of doule MP3 compression [J]. Cognitive Computation, (4: 9-96,. [9] Qiao M, Sung A H, Liu Q. Revealing real quality of doule compressed MP3 audio [C]. Proceedings of the international conference on Multimedia,. [] Bianchi T, De Rosa A, Fontani M, et al. Detection and classification of doule compressed MP3 audio tracs [C]. Proceedings of the first ACM worshop on Information hiding and multimedia security, 3. [] Yu R, Lin X, Rahardja S, et al. A statistics study of the MDCT coefficient distriution for audio [C]. In ICME, 4. [] 454