Digital video steganalysis by subtractive prediction error adjacency matrix

Video has become an important cover for steganography for its large volume. There are two main categories among existing methods for detecting steganography which embeds in the spatial domain of videos. One category focuses on the spatial redundancy and the other one mainly focuses on the temporal redundancy. This paper presents a novel method which considers both the spatial and the temporal redundancy for video steganalysis. Firstly, model of spread spectrum steganography is provided. PEF (Prediction Error Frame) is then chosen to suppress the temporal redundancy of the video content. Differential filtering between adjacent samples in PEFs is employed to further suppress the spatial redundancy. Finally, Dependencies between adjacent samples in a PEF are modeled by a first-order Markov chain, and subsets of the empirical matrices are then employed as features for a steganalyzer with classifier of SVM (Support Vector Machine). Experimental results demonstrate that for uncompressed videos, the novel features perform better than previous video steganalytic works, and similar to the well-known SPAM (Subtractive Pixel Adjacency Model) features which are originally designed for image steganalysis. For videos compressed with distortion, the novel features perform better than other features tested.