Research On Deepfake Videos Detection Method Based On Deep Learning

With the rapid development of artificial intelligence technology,the spread of deepfake videos on the Internet has posed potential threats to the country,society and individuals.How to quickly and accurately detect deepfake videos from massive videos has become a new key task in cyberspace security.In previous detection methods,there are some problems such as insufficient facial feature extraction and low detection accuracy,which can not effectively balance the detection time and the detection accuracy.This thesis proposes two new methods for deepfake videos detection based on deep learning to further improve the detection effect.The main contents of this thesis are as follows:1.A preprocessing scheme for deepfake videos is proposed.Video data is converted into image data by video frame extraction and face detection;At the same time,three data augmentation methods based on information deletion are proposed to enhance the generalization ability of the model while solving the problem of insufficient sample size of the existing deepfake videos datasets.2.A deepfake videos detection method based on i Res Net34 model and data augmentation is proposed,which names i Res Net34-DA.i Res Net34 model firstly uses group convolution instead of ordinary convolution to extract more sufficient facial features of forged faces without increasing model parameters.Then the shortcut branch of the dotted residual structure was optimized,using the maximum pooling layer to replace the convolution kernel to complete the downsampling operation,so as to reduce the loss of facial feature information.Then,the channel attention layer is introduced after the convolution layer to increase the weight of the channel which extracts the facial key features.Finally,i Res Net34 model is used to train datasets which are expanded by three data augmentation methods based on information deletion,and the detection effect of deepfake videos is improved.3.A deepfake videos detection method based on capsule network and spatial-temporal feature fusion is proposed,which names Caps Net-i DRA-STF.The method uses capsule network to further extract face spatial features,and proposes i DRA algorithm by optimizing the dynamic routing algorithm.Dynamic routing update is regarded as a similar clustering process,using cosine similarity to optimize initial weights of low-rise capsule.Then,in the iterative updating process of the dynamic routing algorithm,the weight of the noise capsule whose dot product is negative with the initial clustering center is given zero,so the model has stronger ability to represent forged faces.At the same time,the optical flow algorithm is used to extract the inter-frame temporal features of the deepfake videos,and the model is used to fully learn the spatial and temporal features of the deepfake videos,so as to further improve the detection effect.4.The experimental analysis and comparison of the two detection methods are carried out.The results show that with the strategy of using convolution neural network and extracting only spatial features,the detection accuracy of i Res Net34-DA method on Face Swap and Deepfakes datasets is 99.33% and 98.67%,but the detection effect is reduced on Celeb-DF dataset with higher forgery quality.With the advantages of using capsule network and spatial-temporal feature fusion,the detection accuracy of Caps Net-i DRA-STF method on Celeb-DF dataset is94.07%,which is significantly higher than i Res Net34-DA method.At the same time,it is concluded that i Res Net34-DA method has higher detection efficiency while Caps Net-i DRASTF method has higher detection accuracy.In the actual detection task,the appropriate detection method should be selected according to the specific situation.