Double H.264 Video Compression Detection Based On Abnormal Pattern Of Prediction Residual Distribution

In recent years,the development of multimedia technology is changing rapidly.People can watch digital video through different ways now.Meanwhile,criminals can use the digital video editing software to forgery the videos,which may trigger public opinion,even make perjury.Thus,video forensics technology is getting more and more attention by researchers.Since the double compression must occur in a video forgery process,video double compression detection is of vital importance for the video forensics.This paper focuses on the GOP(Group of Pictures)not-aligned H.264 double compression detection.By using the PRED(Prediction Residual Distribution)feature of the H.264 video.Two H.264 video double compression detection methods are proposed.The main innovations of this paper are listed as follows:1.A novel double H.264 compression detection scheme based on PRED analysis is proposed.For each frame of a given video,the PRED feature is first extracted.The JSD(Jensen-Shannon Divergence)is introduced to measure the difference between the PRED features of adjacent two frames.After that,the denoising method is used for the JSD sequence.Finally,a Periodic Analysis method is applied to the final feature sequence to detect double H.264 compression and to estimate the first GOP size.Experiments have demonstrated that the proposed scheme can achieve better performance than the state-of-the-art method investigated.The GOP estimate accuracy is also more accurate.2.A novel double H.264 compression detection scheme based on multiple features is proposed.Both PRED feature and VPF(Variation of Prediction Footprint)feature are used to generate the feature model.By using the frame-level prediction theory,all the P frames of the video are classified through SVM(Support Vector Machines).Then the proposed method will analyze the prediction probability sequence to detect double H.264 compression as well as to estimate the first GOP size.The experiments show that this method can achieve better performance compared to the PRED method.