Research On Video Motion Vector Steganalysis Based On Deep Learning For H.265/HEVC

Digital steganalysis investigates the problems of how to identify stego-media containing secret messages and how to extract the embedded secret messages.As an important research topic in multimedia information security,digital steganalysis has become vital for maintaining national security and social stability.Digital video is a favorable type of cover for steganography because its redundancy in spatial and temporal domains provides high embedding capacity.Steganalysis for digital videos has therefore great significance in theory and practical applications.Unlike image steganography,video steganography almost does not directly modify the spatial pixels of the frame image.Its mainstream embedding method is to modify the coding parameters or DCT coefficient residuals after video coding and compression.Therefore,the signal distortion caused by the modification of the video by embedding is dispersed to the current frame and adjacent frames,This unique property makes it difficult for the existing image steganalysis methods to work directly in video steganalysis.In recent years,deep learning technology has developed rapidly.However,due to the inability to directly sample the embedded subjects,it is difficult for deep learning networks to be used for video steganalysis.As an example,the motion vector(MV)based video steganography embeds messages by modifying the MVs.However,the numbers and locations of MVs vary from frame to frame,making it unable to use the MVs to construct input data matrices of unified size and format.The video steganalysis data tend to be sparsely and unevenly distributed across frames,and lack spatial correspondence between difference frames,which is the major challenge for applying deep learning network.Aiming to solve this problem,this thesis proposes a comprehensive solution to sample detection data from videos and construct input matrices for steganalytic deep learning networks.As a pilot scheme,we design a deep learning network for H.265/HEVC video steganalysis.The major contributions of this thesis are summarized as follows.(1)A data sampling and input matrix construction strategy for detecting video motion vector steganography using deep convolution neural network is proposed for the first time.According to the characteristics of H.265/HEVC video coding,4 × 4-size block as the basic unit,each frame is divided uniformly to solve the problem of uneven distribution of sampling signals;The horizontal and vertical components of the motion vector are sampled simultaneously to solve the problem of insufficient data of the sampled signal;The reference index of motion vector is taken as the sampling object to reflect the correlation between adjacent frames in time domain;According to the different steganalysis contributions of video motion vectors and prediction residuals at different bit rates,the motion vectors and prediction residuals are combined into two branch inputs to complement each other,so as to improve the steganalysis robustness of video at different bit rates.Finally,based on the baseline network SRNet of image steganalysis,a deep convolution neural network VSRNet(Video Steganalysis Residual Network)for detecting H.265/HEVC video motion vector steganography is constructed,and its effectiveness is proved by experiments.(2)A video steganalysis method based on deep convolution neural network with selective channel sensing is proposed.Considering that the probability that the encryption elements are modified in the process of encryption can be used as important reference information for steganalysis,according to the characteristics of H.265/HEVC video motion vector steganography,the embedding probability of motion vector is defined as the sampling object and the corresponding embedding probability graph is constructed to obtain a new sampling data input matrix.At the back end of VSRNet,a selective channel sensing module integrating the information of embedded probability graph is designed to obtain a network SCA-VSRNet(Selection-Channel-Aware-VSRNet)with selective channel sensing.Experiments show that this method can further improve the detection performance of VSRNet.(3)Besides detecting the presence of motion vector modification,it is often required to quantitatively estimate how the modification is implemented.To this end,a deep learningbased quantitative steganalysis method for H.265/HEVC motion vector embedding is proposed,aiming to estimate the embedding rate.The proposed VSRNet is used as the backbone network.Samples with specific embedding rates are used to train the module to extract embedding rate-sensitive features.The features are combined and fed into the embedding rate estimation module for regression.Experimental results demonstrate that the proposed network Q-VSRNet(Quantitative-VSRNet)can accurately estimate the embedding rate for H.265/HEVC video motion vector embedding.(4)We propose a video steganalysis method based on deep convolution neural network combining motion vector difference(MVD)sampled data.On the basis of VSRNet,the strategy of data sampling and input matrix construction is extended to MVD sampling data,and a multi-channel input IVSRNet(improved vsrnet)combining motion vectors,prediction residuals and MVD is obtained.Experiments show that IVSRNet can effectively improve the steganography detection performance of existing VSRNet,SCA-VSRNet and Q-VSRNet.This fully proves that the data sampling and input matrix construction strategy proposed in this paper has good universality and scalability.