Digital video steganalysis based on motion vector statistical characteristics

Unlike traditional video steganography, motion-vectors-based steganographic approaches embed messages by modifying the motion vectors of the cover video. The scheme may produce little quality degradation of the stego video and little influence on the statistical characteristics of the spatial or frequency coefficients of the frames. As a result, the existing video steganalytic algorithms based on the statistical features of frames or the correlation between adjacent frames cannot effectively detect the motion-vectors-based steganographic system. In this paper, an improved steganalytic method against motion-vectors-based steganography is proposed. Based on the proposed algorithm, the correlations between motion vectors both in spatial and temporal domain are effectively quantified and a novel 12 dimensional statistical feature vector is extracted. The support vector machine (SVM) is trained with these vectors to detect the presence of the hidden data. Compared with other algorithms, the proposed scheme has higher detection accuracy for the typical motion-vectors-based steganographic algorithms.