Research On Digital Image Steganography And Steganalysis

With the rapid development of modern computer technology and network communication technology, information security has attracted more and more attention. As two important branches of information security, steganography and steganalysis have experienced more than ten years of development. Steganography and steganalysis play an important role in areas of business, intelligence, national security and so on. In this dissertation, we have proposed one steganography method and three steganalysis methods including a special steganalysis of LSB matching, a universal spatial domain image steganalysis and a universal JPEG image steganalysis. The contributions of this dissertation are summarized as follows:1. Classical LSB(least significant bit) matching steganography embeds message in a single pixel, not considering the texture distribution of natural image. The texture regions of natural images take on complex structural features. Compared with flat regions, image textures exhibit more randomness in structure. As a result, embedding secret messages in rich textures would be more secure. We proposed a pixel block based adaptive steganography algorithm that gives higher priority to texture regions when embedding secret messages. We gave the criterion for assessing textural blocks which is adaptive to the length of the secret message. To deal with possible irregular embedding blocks, we proposed a scheme to modify the pixel values. The scheme is mathematically justified. Experimental results have demonstrated that our algorithm outperforms two established LSB embedding algorithms and one recent edge adaptive steganography algorithm in terms of embedding efficiency. Meanwhile, our algorithm also demonstrates stronger resistance to three representative universal spatial domain image steganalysis algorithms, three representative DWT domain image steganalysis algorithms and one special steganalysis algorithms.2. Having analyzed the impact of LSB matching embedding on the grayscale histogram, we proposed an improved special steganalysis algorithm based on the histogram characteristic function of center of mass(HCFCOM). We proposed to discard the most important bit plane of a grayscale images and then reconstruct a grayscale image with the rest bit planes. Such operation could eliminate the impact of image content and reduce the number of bins in the histogram. Based on the strong local correlations between adjacent pixels, we proposed to further extend the two-dimensional histogram in Ker’s method to multiple directions. The experimental results show that the detection precisions have been improved significantly compared to two existing special steganalysis methods that are also based on histogram analysis.3. Having analyzed the impact of LSB matching embedding on the difference image, we proposed a universal spatial domain image steganalysis method based on Markov Mesh Models. Features are extracted from a conditional probability matrix described by Markov Mesh Models. Moreover, the extracted features are calibrated in image domain by image calibration technique to improve the detection rate. Extensive experiments show that the proposed scheme can not only achieve high detection rates against the classic LSB matching steganalysis methods but also detect two recent texture adaptive steganography algorithms.4. A universal steganalysis method was proposed by using the superior property of Contourlet transform with representation of an image. Extracted features include the coefficient moments statistics, noise residual moments statistics, and characteristic function moments in the high frequency sub-band of Contourlet domain. A non-linear SVM classifier was exploited for classification. We conducted experiments to attack steganography algorithms like JSteg, Jphide, F5 and Outguess with different embedding rates. Experimental results show that compared with the classical wavelet, Contourlet transform is more effective for detecting the slight difference caused by messages embedding. Moreover, the proposed methods also detect two recent texture adaptive steganography algorithms.