Research On Forensics Aided Steganalysis Of Heterogeneous Images

As one of the important technical means for defending the security of information, image steganalysis has become an attractive research topic in the field of multimedia information security. Existing researches about steganalysis have achieved fruitful results, which exhibit excellent detection performance under the laboratory environment. However, the detection accuracy is not satisfactory when the existing steganalysis algorithms are applied to the heterogeneous images under the actual network environment, which contain a variety of image generation source, a variety of image content and texture, and images undergoing a variety of complex image processing.In this thesis, image forensics technique is used for pre-classifying the heterogeneous images to solve the "mismatch" problem of training and testing samples when existing steganalysis algorithms are applied to the actual network environment. The main research work of this thesis is summarized as follows:1. The statistical properties of the heterogeneous images are analyzed. First, the statistical property of the images with different content complexity is analyzed by the method of texture description, including difference histogram, co-occurrence matrix and wavelet transform coefficients CF and PDF moments. Then, the correlation and periodicity of different resampling images are analyzed. Finally, the statistical properties of different smoothing filtering images are analyzed and the changes in correlations between pixels are described using difference histogram.2. A new image steganalysis method using k-means clustering is presented, targeting at the image with different texture complexity. The image content complexity and "mismatch" between training image database and testing image database have a negative impact on steganalysis. To deal with this issue, a new image steganalysis method using k-means clustering is presented. In the training phase, the input images are classified to several classes using k-means clustering according to the texture complexity, and then the training process is specialized for each class separately. In the testing phase, the given test image is firstly classified to the class it belongs to according to its texture complexity, and then it is submitted to the corresponding steganalysis classifier. This method can reduce the mismatch penalty considerably. The experimental results aiming at LSBM and adaptive steganography demonstrate that the proposed method can significantly enhance the detection accuracy of the existing steganalysis methods in contrast with the traditional steganalysis framework.3. A steganalytic method using multi-classification of resized images is presented to handle the steganalysis problem on the heterogeneous images constructed by single-sampled images and resampled images with different interpolation algorithms and scaling factors. The impact of different interpolation algorithms and scaling factors of resampling on steganalysis, as well as the impact of "mismatch" on steganalysis is analyzed firstly in this paper. Then a multi-classifier based on SVM is constructed to classify resized images into multiple categories which are not affected by steganographic embedding operation basically. Finally, a steganalytic method using multi-classification of resized images is presented. This method can reduce the mismatch penalty considerably and improve the detection accuracy of the existing steganalytic methods under practical network environment. Experimental results validate the availability of the proposed method.4. An image smoothing filtering forensics aided steganalyzer of heterogeneous images including raw images and smoothing filtered images is proposed. The impact of different smoothing filtering operation on steganalysis, as well as the impact of "mismatch" on steganalysis is analyzed firstly in this paper. Then an image smoothing filtering detection method using local linear transform is proposed to distinguish smoothing filtered images and raw images, and can identify the smoothing filtering method as well. Lastly, a steganalysis method using image smoothing filtering forensics is presented, which can reduce the mismatch penalty and increase the detection accuracy of the existing steganalytic methods under practical network environment. Experimental results validate the availability of the proposed method by taking LSBM steganography for example.Finally, the research work in this thesis is summarized and the further research work is prospected.