Deepfake Video Detection Algorithm Based On Keyframe

Deepfake poses a great threat to the credibility,reliability and security of information.Criminals use face swap to generate images and videos to mislead the masses,causing adverse effects and even panic.In order to resist the spread of false information,Deepfake detection technologies have attracted extensive attention.To avoid information redundancy and complex calculation,most of the existing video detection methods randomly select multiple frames or segments of video.However,this strategy will reduce the representativeness and limit the performance.In order to solve this problem,we propose a method of using key frames instead of videos as input.On this basis,Deepfake videos are further detected.The research work of this paper is as follows:(1)Based on the problem that the existing algorithms cut the spatial and temporal features of video during detection,a spatio-temporal two-stream network video detection algorithm based on key frames is proposed.In this algorithm,key frames in the videos are selected through inter-frame difference.The hidden forged information in the key frames is fully mined through space and time stream,and then the space and time features are fused and input into the dynamic routing algorithm to obtain the final classification result.Experiments show that the performance of this algorithm in Deepfake video detection and forgery type recognition is better than the existing methods,and the speed is also greatly improved.(2)Based on the obvious performance degradation of existing algorithms in generalization,a video detection algorithm of multi-feature fusion in frequency domain based on key frames is proposed.This algorithm uses the mean square error of frequency domain information to extract key frames,then extracts the artifact features of the main frame and the time inconsistency in the key frames,fuses them and inputs them into the full connection layer to obtain the final detection results.Experiments show that the performance of this algorithm in cross-dataset detection task is better than the latest existing methods,and has strong generalization.