STEGANALYSIS OF IMAGES CREATED IN WAVELET DOMAIN USING QUANTIZATION MODULATION SHAOHUI LIU, HONGXUN YAO, XIAOPENG FAN,WEN GAO Vilab, Computer College, Harbin Institute of Technology, Harbin, China, 150001 E-mail: {shaohl, yhx, fxp}@vilab.hit.edu.cn, wgao@ict.ac.cn The goal of steganography is to avoid drawing suspicion in the transmission of multi-medium hidden message. But some evil may use medium to carry out some uncontrolled and unlimited information The multi-medium industry and many experts have aware of the ponderance of situation in network security and copyright protection. In this paper, we proposed a novel method to reveal existence of hidden communication in stego medium according statistics discrepancies in images, which had been hidden message. 1. Introduction Information hiding has become the focus of research now. This is the art of hiding a message signal in a host signal, such as audio, video, still images and text document without any imperceptible distortion of the host signal. There exist many applications, which are classified mainly into steganography and copyright marking in [1]. Recently years, these two research areas have proposed many schemes and industries have adopted some schemes to add copyright mark into their products. However, accompanying with the positive function of new technology, the negative effect turn up, some tools with these new technology can be abused by evils to pass on illegal message over public network for avoiding law enforcement. And people have waked up to the urgent situation for preventing this tendency. Like cryptanalysis and cryptography, Steganalysis emerges as the times require. Steganalysis is the art of discovering and rendering useless such covert messages, hence making information hiding failed. It is said after 911 terrorists attack, people think terrorists may be communicated with each other by data hiding technology. Tuck Kelley wrote US officials and experts say steganography is the latest method of communication being used by Osaw bin Laden and his associates to out for law enforcement [2]. In this case, the data hiding in digital media on Internet has proven to be a boon for terrorists. So how to determine whether digital media has hidden information become a emergency problem for social security and stability. Some definitions and several methods of steganalysis were proposed in the literature [2~14]. In [3], authors gave an overview of some characteristics to detect the existence of hidden information. And [4] made a good description of popular free software s steganalysis. In [5] authors proposed a steganalysis 1
2 method based fisher linear classifier. The regress analysis in [6] was used to analysis image based image metrics. Hypothesis test in [7] was used to determine whether information exists or not and a general framework was proposed in [8]. And for LSB embedding methods, the most successful researchers may be J. Fridrich[9~11]. Some good summarize papers include [11,12]. And with the help of steganalysis, ones can find more robustness methods to resist attacks and analyses [13]. 2. Energy Difference Based Detection 2.1. Histogram Analysis In current data hiding schemes quantization is an important means to embed message into cover medium. And steganography in transform domain is useful for taking advantages of perceptual criteria in the embedding process. So a large number of methods for hiding messages in raw lossless compressed images (BMP, RAS, PGM, and many other formats) are based on wavelet transform of every gray-scale or color channel with message bits [15~19]. In those methods, most of them use quantization modulation idea. For example [18], people can construct a mean quantizaitonal codebook in implementing that scheme. In [16,17], the quantization process is taken between DWT middle frequencies pair (MFP), which is defined as a pair of coefficients that are at the same location in LH and HL bands of DWT coefficients. In this paper we will focus mainly on these schemes based on quantization in wavelet domain. First we adopt different wavelet domain methods to embed message. Image (a) in Figure 1 is the host image, image (b) use QIM [15] method in wavelet domain, image (c) use HVS and adaptive modulation methods, and image (d) has been hidden messages with method [16,17]. Obviously these images distortions are imperceptible. So we take wavelet transform of images, then getting three sub-bands coefficients, Figure2 shows histogram analysis respectively. (a) (b)
3 (c) (d) Figure1.Hidden message images, where (a) host image,(b),(c),(d) hiding message with method [15],HVS and [16] separately. (a) (b) (c) (d) Figure2. Histogram analysis, where (a),(b),(c),(d) is the histogram of Fig1(a), (b),(c), (d) separately 2.2. Spectrum Analysis and Energy Difference From Figure2, we find that histogram shape of that host image without any message hidden is smoother than images with hidden message. For quantitative analysis of images features with hidden message, we introduce spectrum analysis and energy difference to score these obvious features.
4 First we take Fourier transform images sub-band coefficients. C denote sub-band coefficients, and Curv denotes curve of C, Ĉ denotes fourier transform coefficients, Curv denotes curve of Ĉ. It is called the energy of images that areas of the region surrounded by the curve Curv and the Curveĉ, denoted by Energyc and Energyĉ then e ĉ e c e c Cˆ = FFT( C) Then we search local minimum value in the curve Curv minimum value points are denoted by L min e ĉ (1), these local L min = LocalMin( Curve c ˆ) (2) Curve denotes the curve composed by L min, computing energy Lmin difference between these two curves Curve and Curv e Lmin ĉ EnergyDifference = Energy Energy cˆ Lmin (3) It is deduced that the energy difference for hidden messages images with quantizational modulation method is much higher than no hidden images from the distribution of energy difference curve. And subsequently, many experiments verify this fact. 2.3. Wavelet Filters Selection In the embedding processing, different people may select different wavelet filters and decomposition level, hence we list most of all popularly known wavelet filters to make this detector more widely applicability. And in implementation, selecting two levels to four levels decomposition, we choice the maximum energy difference as the score to determine existence of hidden message in images. 3. Detection Results The data hiding techniques we used were: Brain Chen s quantization index modulation (QIM) methods [15], DWT middle frequency pair (MFP) [16,17] and
ourselves scheme based adaptive quantization method (AQM). One reason for the selection of these techniques was their availability and the fact that they were all known algorithms based quantizational modulation in wavelet domain. We used an image database partly from USC-SIPI s miscellaneous part for the simulations, in addition some from digital camera and Internet. These test images contained a variety of images including computer generated images, images with bright colors, images with reduced and dark colors, images with textures and fine details like lines and edges, and well-known images like Lena, peppers etc. And the test image base includes (no hidden) and (hidden) images. The result shows in Table 1 Table 1. Test results Method Number of images QIM MFP AQM Type of images hidden hidden hidden no hidden Right detection 763 760 762 756 Rate 99.9% 99.4% 99.7% 98.9% From this table, we found that this is an excellent result, for these three methods, theirs successful detective rate reach up 99%, and the false positive alarm is very low. 5 4. Conclusion This paper provides a new method to detect the existence of hiding information. We focus on those methods that adopt quantizational modulation in wavelet domain. Using histogram and spectrum analysis, we get a quantitative detector to those methods. And results indicate that the method is valid. Although result is very exciting, there is a lot of work that still needs to be done. Many other watermarking schemes and algorithm will to be included this research and extensive tests need to be done with a larger number of images. Detection technique development in this area of data hiding will continue. We find image s statistic features are important clues to determine whether hiding information or not from the detection process. So this suggest us to develop more robustness method with statistic features altered as little as possible. References 1. http://www.usatoday.com/life/cyber/tech/2001-02-05-binladen.htm.
6 2. N. F. Johnson and S. Jajodia, IEEE Computer 2, 26(1998). 3. N. F.Johnson and Sushhil Jajodia, Proceedings of the IEEE Information Technology Conference, (1998). 4. N. F. Johnson and S. Jajodia, Lecture Notes in Computer Science. 1525, 273 (1998). 5. H. Farid, Technical Report, TR.01, 412 (2001). 6. I.Aveibas, N.Memon, B.Sankur, 2001 IEEE Fourth Workshop on Multimedia Signal Processing, 517 (2001). 7. R. Chandramouli and N. Memon, Analysis of LSB based Image Steganography Techniques, Proceedings of ICIP, 206 (2001). 8. R.Chandramouli, N.D.Memon, A Distributed Detection Framework for Steganalysis. ACM Multi-media Workshop Marina, 123 (2000). 9. J. Fridrich, R. Du, and L. Meng, Proceedings IEEE International Conference on Multimedia and Expo, 68 (2000). 10. J. Fridrich, M. Goljan, and R. Du, SPIE Multimedia Systems and Applications IV, 275 (2001). 11. Jessica Fridrich, Miroslav Goljan, Proceedings of SPIE. 4675, 1 (2002). 12. N. Provos and Peter Honeyman, CITI Technical Report. 11, 1 (2001). 13. N. Provos, 10th USENIX Security Symposium, (2001). 14. S.Voloshynovskiy, A.Herrigel, Y.Rytsar and T.Pun, Edward J.Delp III, Ping Wah Wong(Editors), Proceedings of SPIE. 4675, 57 (2002). 15. B. Chen and G. W. Wornell. Implementations of quantization index modulation methods for digital watermarking and information embedding of multimedia. J. VLSI. 27, 7 (2001). 16. Quan Yuan,Hongxun Yao,Wen Gao and Sanghyun Joo, Blind Watermarking Method Based on DWT Middle Frequency Pair. Proceeding of IEEE ICME. 2 473 (2002). 17. Quan Yuan,Hongxun Yao,Wen Gao and Sanghyun Joo.Blind watermarking using HVS model and wavelet transform. Submitted to IEEE transaction on Circuits and Systems, (2002). 18. Jim Chou, Sandeep Pradhan and Kannan Ramchandran, A robust blind watermarking scheme based on distributed source coding principles. ACM Multimedia, (2000). 19. Chun-Shien Lu, Hong-Yuan Mark Liao and Chwen-Jye Sze, Combined watermarking for image authentication and protection. IEEE ICME. 3, 1415 (2000).