The Research On Passive Steganalysis For Digital Video

Digital steganography, which is widely used in the field of covert communication,is an important branch of information hiding. To preserve the security ofcommunication system as much as possible, steganography embeds secreteinformation into cover data in an imperceptible way. As an opposite of steganography,steganalysis is a technique for detecting and attacking steganography. Its purpose is toexamine whether there are suspicious messages in cover data or not. With thedevelopment of video steganography and the emergence of video steganographysoftwares, video steganalysis is becoming a hot research topic in the field ofinformation security.In this thesis, the differences of statistical properties between cover-video andstego-video are analyzed, and those feature vectors that can capture those differencesare extracted effectively. Then, support vector machine is utilized for patternclassification and passive steganalysis for digital video is achieved finally. The mainworks are summarized as follows:First, for LSB matching steganography, a video steganalysis is proposed based onthe local area’s correlation. LSB matching steganography is most difficult insteganalysis. After an analysis of the statistical properties’s change in cover-video dueto steganography, an image of regional correlation is constructed. The1D and2Dhistogram feature vectors are extracted, and then median filter is utilized to decreasethe feature differences among different videos. A video steganalysis scheme isimplemented in a frame-by-frame manner by using support vector machine.Experimental results show that the proposed algorithm can detect not only thosevideos in every frame and an minimum embedding strength of0.1, but also for thosevideos with secret messages in only20%of frames and embedded in a mixed way. Adetection accuracy of90.91%is achieved.Second, a video steganalysis algorithm is proposed based on bi-directionalmotion estimation collusion. An estimated frame is obtained by bi-directional motionestimation collusion, and multiple histogram characteristics are extracted fromprediction error frame. The classification of cover-video and stego-video is achievedby support vector machine. Compared with collusion attacks by traditional temporalframes averaging (TFA) and its improvement MC-TFA, the proposed approach highlights the difference of prediction error frame between innocent-video andstego-video. Moreover, it does not need the assumption of Gaussian property ofprediction error frame. The results of simulation experiment demonstrate that forspread-spectrum steganography with an embedding strength of1, the detectionaccuracy of passive steganalysis can be up to96.57%.