Research On Passive Detection Of Deepfake Video Based On Deep Learning

In recent years,Deepfake based on deep learning has developed rapidly.Yet it has also been taken advantages by outlaws to disseminate false information,damage others’ reputation and seek illegal interests,which has seriously endangered social stability and national security.The existing Deepfake detection methods generally have some deficiencies,such as inaccurate face location,long detection time,large resource consumption,low accuracy of authentication and poor visualization effect.Aiming at the above problems,taking Deepfake videos as the research object,a passive Deepfake detection system based on deep learning is studied.In view of the inaccuracy of face location in Deepfake detection,this thesis improves the Retinaface face detection network,replaces the baseline network model with the MobileNet-V2 lightweight network and builds a face detection module in combination with feature pyramids and enhanced feature extraction layers.The preprocessed face detection dataset is used to train and test the improved Retinaface face detection network.Subsequently,the improved network can improve the detection time and accuracy.In view of the long detection time,high resource consumption and low accuracy of the Deepfake detection,this paper proposes for the first time the use of EfficientNetV2 network for fake video identification,and the network model is scaled appropriately.The Efficient Channel Attention module is used to improve the basic modules in the EfficientNet-V2 network.This thesis preprocesses Faceforensics++ data set and Face Forensics in the Wild through video framing method.The improved EfficientNetV2 network is trained and tested using the preprocessed image data set,which effectively improves the training speed and accuracy of the authenticity verification network.In view of face misdetection and missed detection in Deepfake video detection,this thesis builds a dual-network detection framework,which connects the face detection network and the authenticity identification network,so as to realize face detection and then authenticity identification.This paper uses the video frame test set to test the performance of the detection framework,and proves it reduces the false detection rate and missed detection rate of faces in videos.Moreover,due to the problem of poor visualization effect of Deepfake detection technology,this paper develops a Graphical User Interface(GUI)visualization system to carry out the detection of real and fake pictures and videos.Through actual testing of pictures and videos,the GUI system could effectively optimize the visualization of output results.