Research On Video Tampering Detection Based On Deep Learning

With the rapid development of Internet technology and multimedia digital technology,the application of digital video becomes more and more extensive.A large number of powerful video tampering software has made a great challenge to the integrity and authenticity of digital videos,and therefore the video forensics technologies were proposed to protect to the integrity and authenticity of digital videos.The passive tampering forensics technology of digital videos have become an important solution to protect the integrity and authenticity of videos.In traditional digital video forensics techniques,the tampering features need to be extracted by professionals,and complex calculations and appropriate thresholds are used to determine whether or not a video has been tampered,resulting in low accuracy and complicated calculations.In recent years,deep learning methods have made great breakthroughs in the field of image recognition,providing a new idea for digital video forensics technology.Based on the analysis of video tampering and video compression coding,two video tampering detection algorithms based on deep learning are proposed in this paper.Experiment results show that our video tampering detection algorithms have the characteristics of easy operation,high accuracy and fast speed.The main works of this paper are described as follows:(1)An video tampering annotation technique based on regression algorithm is proposed in this paper.First,in our technique,a manual interval labeling method is used to label a small part of video frames so as to obtain video annotation data.Then,a regression model that matches the distribution characteristics of the obtained annotation data is selected.Next,the annotation data are used to adjust gradually the structural parameters of the mode,so that the trained regression model with the smallest error loss function is obtained.Finally,the obtained regression model is used to predict the remaining unlabeled video frames in order to get all the annotation data of the video.The method greatly reduces the cost of manual labeling,and at the same time speeds up the efficiency of labeling,and is suitable for large-scale data labeling tasks.(2)Two video tampering detection algorithms based on deep learning is proposed in this paper.One is the video tampering detection algorithm based on residual network.In this algorithm,the residual network is used to increase the number of network convolution layers,which improves the network feature extraction and nonlinear fitting ability,and further improves the detection accuracy.The experimental results also show that the recognitionaccuracy of the network is significantly improved compared with the traditional algorithms.However,this algorithm has the disadvantages of large computational complexity,low operational efficiency,and incomplete extraction features.In order to resolve the above shortcomings,we propose the second algorithm,a video tampering detection algorithm based on Google Net's two-stage classifier.In the second algorithm,the residual images of video frames are first extracted by using the first-stage classifier,and then the high-pass filter and the collusion operation are used to enhance the artifact information introduced by the tampering operation in the residual images,and finally the threshold classifier is used to classify video sequences into tampering videos and original videos.In the second-stage classifier,Google Net's Inception-V3 network is used to locate tampering video frames.Specifically,the network extracts the deep features of video frames through a variety of convolution kernels so as to detect whether or not the video frame has been tampered,and locate the locations of the tampering video frames in the video sequences.Besides the high recognition accuracy,the second algorithm has localization ability,easy processing,low time cost and practicability.