| In recent years,Deep Fake technology has developed rapidly.The face forgery technology achieves the visual effects of face identity replacement and expression replacement by tampering with real face images.Some criminals with unruly schemes attempt to achieve illegal purposes by this,which brings risks to politics,economy,and social security.With the advancement of forgery technology,the forgery process has shifted from the replacement of the entire face to the more concealed local tampering,which brings the difficulty of local information localization to the face forgery detection.In order to solve the problem of local forgery detection difficulties,strengthen the use of image-level supervision labels that can be obtained in real scenes,start with the design of the network structure framework and loss function,and enhance the local forgery detection capabilities of the model.The specific research content includes:Firstly,based on the problem of too large receptive field of neural network classifier and high similarity of local features,design a Local-Prediction framework and a Local-Diversity loss under restricted receptive fields.The Local-Prediction framework based on the restricted receptive field adapts the receptive field range of the local forgery features,focusing on the detection of local forgery;the Local-Diversity loss provides regularization terms and enhances the difference between adjacent local features to enrich the learned characteristic pattern.Based on the results of local prediction,the final prediction focus on pixels with relatively high activation values.By limiting the receptive field and setting the rules of diverse features,finally formed a forgery detection framework focusing on local feature learning.Secondly,based on the relatively single problem of local feature mode,design a multi-scale face forgery detection scheme that integrates global and local features.The local branch detect the forgery features of key local areas by limiting the image scale of the input model;the global branch focuses on the difference in the overall forgery pattern,focusing on the noise consistency of the foreground and background before and after tampering.The two branches form a complementary and mutually beneficial multi-scale information extraction process,and finally form a joint judgment through feature fusion,which simulates the multi-scale judgment process of the naked visual system and improves the robustness of forgery detection.Experiments on the benchmark dataset Face Forensics++,which contains a variety of counterfeiting methods,Deep Fake,Face2 Face,Face Swap,and Neural Textures,show the effectiveness and feasibility of the proposed algorithm in this paper.In the experiment,the detection accuracy of Neural Textures,a local forgery method,has improved significantly,indicating the effectiveness of the local forgery detection design.In the test of the Celeb-DF and DFD datasets,the performance of the algorithm is significantly improved compared with the existing binary classification methods. |
