Research On Key Issues Of Detection For Digital Image Steganography

Reliable detection of stego multimedia files has important applications and great theoretical values. Although after more than ten years'study researchers have made a great progress in steganalysis, there are still many key issues needed to be researched and solved. This dissertation mainly focuses on the detection for digital image steganography, and includes eight chapters which are divided into three parts.The first part: background knowledge and state of the art. With respect to the application and technical background, the practical and theoretical value of steganalysis are described. The basic concepts, models and research review of information hiding are introduced briefly. The research progress of image steganalysis is surveyed in detail.The second part:the detection for specific steganography. Based on the fixed change mode of image pixel pairs caused by the replacement steganography, the detection algorithms for LTSB (Least Two Significant Bit) replacement steganography and MLSB (Multiple Least Significant Bit) adaptive steganography of spatial domain are proposed. Based on the relative entropy of DCT coefficient histogram, a detection algorithm for F5 steganography is also presented.1. For LTSB replacement steganography of spatial domain, by analyzing the sets of adjacent pixel pairs of an image before and after message hiding, some statistical laws of natural images are found. On this basis, based on the principle of similar triangles, an algorithm for estimating embedding ratios of LTSB replacement steganography is proposed. Experiments show that the proposed algorithm can estimate the embedding ratio of LTSB steganography accurately, and overcome the ill-conditioned problems of the existing methods in the context of high embedding ratios.2 For MLSB adaptive steganography of spatial domain, a general detection method is presented based on specific regions and pixel pairs analysis. Considering the non-flat areas as particular regions, an algorithm for selecting specific areas and an algorithm for estimating embedding ratios of MBPIS adaptive steganography are given. Experiments show that the proposed algorithm is able to estimate the embedding ratios on each plane of image pixels more accurately.3. For F5 steganography of frequency domain, first, estimate the original image by calibrating the test image, and calculate the coefficient histogram of stego image with embedding ratio p; second, based on the relative entropy of DCT coefficient histograms of the test and stego image, the change ratio of image DCT coefficient is obtained, which can be considered as the change ratio estimation of the test image; last, according to the change ratio and the test image, amend the histogram of original image, and give the estimation algorithm for embedding ratios. Experimental results show that compared to current typical detection algorithms, the proposed new algorithm has higher estimation accuracy.The third part:the universal blind detection algorithm and feature comparison. This part focuses on two aspects:the design of steganalysis algorithms and the feature comparison analysis for blind steganalysis. In the first aspect, from the perspective of multi-domain features, several blind detection algorithms for classifying the original images, PS images and stego images are proposed. From the perspective of time-frequency analysis, some blind detection algorithms are given based on full wavelet packet decomposition and best wavelet packet decomposition respectively. In the second aspect, for several wavelet coefficient subbands, the PDF (Probability Density Function) moments and the CF (Character Function) moments are compared, and the theoretical basis of the comparison results are presented.1. Universal blind detection based on multi-domain feature integration. First, the detection framework for classifying the original images, PS images and stego image is proposed; second, three kinds of detection algorithms are given, which include the blind detection algorithm based on the features extracted from image pixels, DCT coefficients and the wavelet coefficient subbands, the classification algorithm of PS images and stego images, and the classification algorithm of LSB replacement and LSB matching stego images. Experiments show that the framework improves the practicality of steganalysis, and the proposed algorithms have higher accuracy and universality.2. Universal blind detection based on optimization time-frequency analysis. From the point of view of optimization time-frequency analysis, the construction method of the features extraction source is discussed. Using the full wavelet packet decomposition, which is neither redundant nor omission, to make the time-frequency analysis for images, the blind detection algorithm based on the full wavelet packet decomposition is proposed. Base on the different best wavelet packet decomposition structure, by adaptively selecting the training features, three different blind detection algorithms based on best wavelet packet decomposition are given. Experimental results show that comparing with the typical existing blind detection algorithms, the new algorithm has a higher correct detection rate and a better universality.3. Feature comparison and analysis. Starting from the experimental results of existing literatures, which indicated that the PDF moment is better than the CF moment with respect to log prediction error subband of wavelet coefficient, the trends of the change of PDF moments and CF moments before and after message embedding are analyzed with respect to prediction subband of wavelet coefficient, prediction error subband of wavelet coefficient, wavelet coefficient subband of noise, and log prediction error subband of wavelet coefficient. The theoretical basis of the conclusion that for the first three subbands the CF moment is better than the PDF moment is presented. Based on the prerequisite that the feature extraction source satisfies Gaussian distribution with 0 mean before and after message embedding, a general conclusion that the CF moment is better than the PDF moment is given; For comparing the PDF moment and the CF moment with respect to the log prediction error subband of wavelet coefficient, the expressions of the CF and PDF moments are given. As the expression of the PDF moment is Lerch transcendental function and it is hard to analysis quantitatively, based on the property of Lerch transcendental function that it can be approximated with any degree of precision, a conclusion that the first order PDF moment is better than the first order CF moment under the case of finite precision is proposed. Finally, a conclusion with a discussion of the direction for the future research is given.