Research On Steganalysis For Jpeg Images Based On Dct Domain

JPEG images are widely used in Internet. They are favorable cover images because of discrete cosine transform(DCT) in JPEG compress standard.This thesis is mainly focusing on steganalysis in JPEG images. On the basis of studies on steganography and steganalysis in JPEG images, the main work and contributions of this thesis are summarized as follows:1,The history, investigation status and application areas of information hiding are introduced. The model, classification and technique research of steganography and the importance, classification and technique research of steganalysis are presented too.2,Based on studying on three classical algorithms(Jsteg, F5, MB), the principles of steganography has been discussed. We have compared them from algorithm performance and the effect on cover images'statistical characteristics, and some countermeasures are introduced. Experiments are done to test these steganalysis algorithms, and the algorithms'shortcomings are discussed.3,A steganalysis for F5 algorithm is proposed based on multi-characteristics. Cover images are easily polluted by F5 steganography algorithm. Based on studying on the principle of the F5 algorithm, 35 characteristics are extracted from DCT histogram, the run-length of zeros and the smoothness of boundary. We use support vector machine (SVM) to construct a set of binary classifier. Experiments are done in many different conditions, and results show that the algorithm keeps a high detecting rate, and it works well. Its most strongpoint is that it improves the detecting performance for low rate stego images.4,A universal steganalysis based on support vector machine (SVM) is proposed. Base on doing research on general principle of steganography in DCT domain and the DCT coefficient statistical model, 46 characteristics are extracted from spatial and DCT domain like the differences in DCT histogram, the change of the mess of DCT histogram, co-occurrence matrix and the smoothness of boundary etc. These characteristics are used to classified cover images from stego images. In experimentation section, Jsteg, F5 and MB stego images are input to test the performance of the algorithm. Experimental results show that the algorithm keeps a high detecting rate in many train conditions for different embedding rate stego images, and it keeps a low false alarm rates.Finally, we summarize our research work for the thesis, and discuss further research topics and directions in the future.