Steganalysis Based On Multimode Texture Classification

As the information security is being focused, more an more countries are researching the method of steganography and steganalysis. JPEG image is widely concerned because of its big amount of number transporting on the internet and its good cryptic performance. Thus, lots of algorithms of steganography and steganalysis of JPEG image have been developed.This paper mainly focused on general steganalysis of JPEG image. We did an intensive study of steganalysis based on multimode texture classification, and implemented a high speed feature extraction mehod with GPU parallel computing. We found the difference of the sensitivities embedding secret message of different texture, and classified with different texture.The main work is accomplished in the thesis as following:1. We came up with texture classification before steganalysis. A nature image has lots of texture type. Different texture has different sensitivies of secret messge. Texture classification could reduce the range of object which is waiting for being detected. We confirmed a texture type by its autocorrelation function and energy, contrast, entropy, correlation of gray level co-occurrence matrix. We classified the texture type by flat and rough, and analysed their characteristic in space and frequncy domain. The result showed that the secret message characteristic of rough texture is more sensitive than it of flat texture in frequency domain.2. Detect different type of texture using OC-SVM(one-class support vector machine). The benefit of OC-SVM is that we could train the classifier just with cover image. This meets the practical use. After the texture classification, the problem of wild range of training sample had been solved, and the classification had a greater performance.3. We took distribution characteristic of dct coefficient for frequency domain feature, extracted multi-direction neighbour coefficient and block edge coefficient, and got a good detection rate.4. GPU parallel comupting. The parallel computing increased the speed of feature extraction for decade times. We took CUDA library produced by NVIDIA.