Research On Steganography And Steganalysis With New Characteristics Of Motion Vectors For HEVC Video

Digital steganography and digital steganalysis,as two important branches in the research of information security,play an important role in maintaining national security and social stability.Among the various media types for steganography,digital videos have become a hot recently.The advanced compression technology together with the efficient network transmission have made digital videos one of the most important information carriers.Compared with digital images,digital videos provide higher embedding capacity due to its higher redundancy in spatial domain and temporal domain.Current research of video steganography and steganalysis is mainly oriented to H.264 or earlier formats.High Efficiency Video Coding(H.265/HEVC)is the new generation of video coding standard which has been increasingly popular.To achieve a better encoding performance,HEVC introduces some new techniques of prediction and coding.While these new techniques facilitate new embedding strategies for steganography,they also bring new challenges for steganalysis.In view of this,this thesis investigates the security of HEVC video from both the steganography and the steganalysis perspectives.In particular,we focus on the advanced motion vector prediction technique of HEVC and propose two types of steganalytic approaches as well as one non-degraded steganographic algorithm.The main contributions of the thesis are summarized as follows:Firstly,current steganalytic methods cannot detect those steganographic algorithms specific to HEVC format.For example,the advanced motion vector prediction technology of HEVC provides a new way of motion vector(MV)modification to embed secret messages.By modifying the prediction indices of the candidate list,the motion vectors can remain unchanged in the reconstructed stego video.Such modification cannot be detected by current steganalytic methods.To solve this problem,we proposed a video steganalytic approach using local optimality in the candidate list.We first exploited Lagrangian cost function to measure the local optimality of the predicted MV in the candidate list.This type of local optimality can capture the disturbance changes caused by different types of MV-based steganography.To improve the effectiveness of the steganalytic method,we combined this type of local optimality with the local optimality of MV,and designed a 40-dimensional feature set in the rate-distortion sense.Experimental results show that compared with three classic steganalyzers,our approach improves the performance against five typical MV-based steganographic methods under different bitrates and embedding rate settings.In particular,the proposed approach significantly boosts the detection accuracy of the steganographic approach based on prediction index modification.Secondly,we proposed to employ the distribution of optimal embedding probability as the prior knowledge for video steganalysis,which is motivated by the fact that most steganography approaches follow the rules of minimum distortion or maximum entropy.The distortion function is defined with respect to the motion features,the textures,and the local optimality under HEVC coding framework.According to the defined distortion function,the embedding probabilities are estimated with Gibbs distribution.We proposed a way of quantitatively enhancing the steganalytic features with the estimated embedding probabilities.The justification of the enhancement is given from the perspective of relative entropy.Experimental results on four classical steganalytic methods demonstrate that by enhancing the features using the proposed approach,the detection accuracies have been unanimously improved while showing robustness against different bitrates.Thirdly,current video steganographic approaches would inevitably incur the quality degradation of stego videos.Utilizing the advanced motion vector prediction technology of HEVC,we explored a new embedding space,i.e.the prediction indices of the candidate lists,and proposed a non-degraded HEVC adaptive steganographic approach.The framework of distortion-minimization steganography with STC is applied to the embedding space of the prediction index.To reduce the disturbance degree of the local optimality of predicted MV in the candidate list,we designed a distortion function using the rate-distortion difference under the prediction of candidate MVs.Experimental results not only prove the non-degrading property of the proposed steganographic approach,but also demonstrate the superiority over the current HEVC steganographic methods in terms of resistance to steganalysis.