Research On Blind Steganalysis For Compressed Video Stream

Covert communication is an important branch of information hiding, which includes steganography and steganalysis. Different from classical cryptography techniques, steganography embeds secrete information into host signal, and try to avoid the change of visual perception and statistical properties as possible, so as to make it safe for transmitting covert message via public channe. As an opposite of steganography, steganalysis is a process of detecting the presence of covert data. Because of the temporal masking effects of human visual system (HVS), digital video provides fertile ground for embedding higher volumes of covert data. Recently, video steganography and its counterpart have gained increasing interests.Though an individual frame may be treated as a still image, existing image steganalysis algorithms will be sub-optimal if they are directly applied to video. From a steganalyst's point of view, they neglect the temporal redundancy inherently existed in videos, which provides a great chance of statistical detection. Compared with image steganography and steganalysis, not much is investigated in the field of digital video. At present, video steganalysis is still far from mature. Most practical video steganalysis methods to date are designed to be passive. In this thesis, we researches on the steganalysis techniques for those collusion-irresistant steganographic algorithms. The main work of this thesis is as follows:Firstly, a novel blind steganalysis framework is proposed for compressed video, which is based on Temporal Frames Weighed Averaging (TFWA) and noise classification. For TFWA based collusion, it allocates different weighting coefficients for the frame within the collusion in terms of its distance to the window center based on first order Markov model. Compared with classical termporal frame averaging (TFA), TFWA can release the influence of those samples with great difference to the collusion results, and restrain the collusion noise. Experimental results demonstrate that TFWA has strong collusion capability. The colluded frame can approach the original frame well by appropriately extending the window redius. The decrease of collusion noise will do benefit to the increase of detection sensitivity to secrete signal.Secondly, an adaptive bi-level noise classification algorithm is proposed for video steganalysis. There are always noise interferences in collusor. Two-level noise classification will divide these noises into different components. It makes use of the easily available information from video stream due to the video coding mechanism, and captures the interferent components caused by inherent change of movement and illumination in videos. Meanwhile, according to the statistical properties of collusion results; those imposed unneglectable components are found. By the definition of content change factor (CCF) of video content within the collusion window, the thresholds for noise classification are adaptively adjusted.Several classical steganographic algorithms, including spatial LSB, F5 and even these two methods associating with spread-spectrum (SS) algorithm with different embedding intensity are utilized for performance test. The proposed video steganalysis algorithm can achieve an accuracy more than 99.8%, when embedding intensity is about 10%.