Research On Video Tampering Detection Algorithms Based On Deep Learning

The development of digital multimedia technology has helped us overcome many obstacles in society,but it has also brought serious threats related to information.With the popularity of various video editing tools and software,ordinary people can easily forge and edit videos.For some illegal elements,tampering with the original video information by using editing software destroys the authenticity and integrity of the video,thus achieving the illegal purpose,and may cause adverse consequences.In order to solve the security problem of video,video tampering detection technology is proposed.Video tamper detection technology has become a research hotspot of domestic and foreign experts in the field of multimedia security.This paper mainly focuses on the application of in-depth learning in video tamper detection technology.The main research work is as follows:(1)Intra-frame tampering detection algorithm based on convolutional neural network and perceptual hash learning.The main idea of the research is to preprocess the video to get the corresponding time-domain representative frame,and then use the time-domain representative frame as the input of convolution neural network to learn the visual characteristics of the spacetime domain.At the same time,the hash layer is designed to learn independent hash functions for each input time domain representative frame and claim the final video hash.Finally,in the optimization objective function,considering the impact of quantization error on the whole framework,it is added to the overall loss function to obtain a more expressive hash.We analyze the validity and feasibility of this method from two aspects: anti-content preservation operation ability and tampering recognition ability.The experimental results show that the algorithm has good robustness and discrimination,and can detect tampering effectively.(2)Video inter-frame tampering detection algorithm based on 3D convolution neural network.The main idea of the research is that the traditional 2D convolutional neural network can not extract the features in time domain,and the concept of 3D convolutional neural network is introduced to design the neural network for tamper detection between frames.The 3D convolution neural network can process a continuous sequence of video frames at the same time,and the resulting map has the information of multiple video frames.Compared with the existing 3D convolution neural network,we design a pixel difference convolution layer,that is,when convoluting in the first layer,we introduce a pixel difference layer in advance to obtain the pixel difference between the input video frames.The difference value will be used as the input of the 3D convolution layer for convolution calculation.After the subsequent convolution and pooling operations,the feature containing a large amount of time information can be obtained for classification.This feature is more sensitive to the performance of tampering operations between video frames.We have experimented with frame insertion,deletion and duplication in static and dynamic single-shot video.This method can show good classification performance,and is superior to the existing technology in the accuracy of inter-frame tamper detection.