Research On DeepFake Face Video Detection Based On Deep Learning

With the continuous advancement and development of artificial intelligence technology,it is becoming easier and easier to tamper with face images in real videos,and the quality of fake face videos generated by DeepFake is getting better and better,even achieving the effect of real ones.Forging face videos will not only threaten the information security of the public and infringe portrait rights of citizens,but also damage the credibility of the government and cause political trust crises and other issues.Therefore,for face video data that is difficult to distinguish between true and false,it has become an urgent task to detect forgery through deep learning methods,which has strong practical significance and social value.Based on this,this thesis fully investigates the current research status in the field,and designs scientific and reasonable experimental methods to conduct research on fake face video detection based on deep learning.The main work of the thesis includes:1,In the research of DeepFake face video detection algorithm based on Video-level,the thesis proposes a forgery detection algorithm based on model fusion.Aiming at the insufficient ability of existing methods to extract information across frames,a self-attention mechanism is introduced to enhance the feature extraction ability of the network for long-distance information in the time dimension.At the same time,the intra-frame information feature extraction branch is added to the network structure,and the lack of the existing methods that cannot mine the intra-frame time domain information is made up for by the model fusion method.Through scientific comparison experiments and ablation experiments,the superiority of the detection algorithm and the effectiveness of the corresponding innovative modules are proved.2.In the research of Frame-level-based DeepFake face video detection algorithm,aiming at the unique spectral features of GAN generated fake images,the thesis proposes a fake detection algorithm based on feature fusion.A cross-modal feature extraction module is added to the network to perform feature fusion of high-frequency image fingerprint information and RGB domain information,which improves the algorithm’s forgery detection effect on GAN forged video frames.At the same time,the thesis extends the traditional two-category forgery detection task to the multi-category forgery traceability task scenario,and designs reasonable comparative experiments to verify the traceability effect of the algorithm method.3.Based on the demand analysis and feasibility analysis of the forgery detection task,facing the actual application scenario,combined with the two forgery detection algorithms proposed in the paper,the paper designs and implements a Deepfake face video detection and traceability system.The front-end page of the designed system is concise and clear,and the user operation is simple and convenient,which lowers the technical threshold of face video forgery detection.Based on deep learning technology,this thesis proposes innovative algorithms from different angles for DeepFake face video detection scenarios,and proves the effectiveness of the proposed algorithm through a large number of experiments.In addition,the proposed detection algorithm is applied in practice in the form of a system,and the functions of forgery detection and forgery method traceability are realized.Therefore,the research of this thesis has certain theoretical significance and application value.