A Multi-Mode Smoothness Blind Steganalysis Based On Clean-Image And GPU Acceleration

Recently along with the development of the computer and internet technology, people can transfer any multimedia information, such as image, video, audio and text. Meanwhile many free and open-source steganography software appeared in the internet, that makes anyone can communicate secretly. If this technology was taken by the terrorist, then it will be a disaster. So steganalysis is becoming the key technology in national security field. This thesis had summarized the advantage and the disadvantage of the existing steganalysis method, discussed the key point of blind steganalysis, and proposed a new steganalysis method which was based on multi-mode smoothness and clean-image.The main achievement of the thesis as following:1. According to the principal of the blind steganalysis, we established the steganalysis model as a light-noise model, and proposed a new concept to classify the image slices by smoothness rate. Using Gradient Energy and compressed ratio as spatial and frequency domain measure separately.2. Proposed different clean-image features in spatial and frequency domain. In spatial domain we use the Gradient Energy, the difference of the Gradient Energy and Gradient Energy with wavelet calibration and the second-attacking Gradient Energy as features. In frequency domain, we use the one-step Markov transition matrix in four directions as features.3. Using One-Class SVM as the classify machine, building multi Hyper-sphere to contain all of the clean-image, so that image which was outside the multi Hyper-sphere was considered to be the dirty-image. That can make the steganalysis method more sensitive to that unknown steganographys, and simplified the training process.4. During the progress of feature extraction, there was a lot of matrix computing. Considering that we use the GPU to accelerate it. A GPU can run thousands threads at the same time, which can increase the computing speed dramatically.