Research On Deepfake Detection Based On Deep Learning

With the continuous upgrading of deep generative models,deepfake technology has become more and more mature.Deepfake is used by people with ulterior motives,which may pose a certain threat to social stability and cyberspace security.At present,deepfake detection has attracted attention from all walks of life.Some deepfake detection methods have been proposed,but most of them use large datasets and have weaker generalization performance.To solve the above problems,through a large number of literatures reading,we combined the advantages of existing deepfake detection methods,we carry out the following research work:(1)A deepfake detection method based on autoencoder.This method has a relatively simple model structure and it is a lightweight model.Our proposed method can effectively detect images generated by nine methods,and has a certain generalization performance.Firstly,the high-frequency feature of the image is extracted,and the encoder is combined with the attention mechanism module to enhance the ability of the model to extract effective features.Secondly,the training efficiency is improved through reconstruction loss and cross entropy loss.Finally,it is the proved by the ablation experiments that the proposed preprocessing method and the addition of attention mechanism module are helpful for the deepfake image detection,and the detection effect and generalization performance of the method are far superior to those of the comparison methods through comparative experiments and generalization experiments.(2)A deepfake detection method using multiple feature fusion.Firstly,the image is preprocessed to extract the high frequency features in spatial and frequency domain.Then,two autoencoder networks are respectively used to extract the features of two preprocessed images,and the extracted two parts of features are fused.Finally,the fused features are input into a classifier that is composed of three fully connected layers through a max-pooling layer.The experimental results show that fused feature input is better than any one kind of these features,compared with other methods,our method is far better than other methods in accuracy and generalization performance.