Research On Generative Face Detection Algorithm Based On Deep Neural Networ

With the rapid development of deep learning technology,high-quality fake face images or videos can be easily generated by using Generative Adversarial Network(GAN)or Deepfake software.The emergence of realistic fake faces has brought great challenges to criminal investigation,justice,and reputation protection,etc.Therefore,there is an urgent need for reliable fake face detection algorithm.Regarding GAN-generated faces,the existing detection algorithms have achieved high indataset detection accuracy,but their generalization abilities need to be improved.Regarding Deepfake videos,the existing detection algorithms cannot provide acceptable in-dataset detection accuracy.This thesis will focus on these two issues.The main works include the following two parts:(1)A novel GAN-generated face detection algorithm with strong generalization ability based on quaternions frequency fingerprint.This algorithm consists of GAN noise fingerprint extraction module and classification module.The former uses quaternion discrete Fourier transform and siamese multiple stages GAN fingerprint extraction network to reveal and extract common fingerprint features respectively.The latter adopts quaternion Res Net to classify the extracted fingerprint features.Experiments based on Celeb A,a popular public facial image dataset,verified the generalization ability of the proposed algorithm.(2)A novel Spatiotemporal Attention based Deepfake detection algorithm.This thesis proposes two novel algorithms named STA-Xcep-Conv LSTM and STA-Xcep-LSTM based on a novel Spatiotemporal Attention mechanism and convolutional long short-term memory(Conv LSTM).The Spatiotemporal Attention mechanism is proposed to capture and enhance spatiotemporal correlations before dimension reduction by Xception.Thereafter,LSTM and Conv LSTM are used to model temporal features in STA-Xcep-LSTM and STA-XcepConv LSTM,respectively.Conv LSTM is introduced to consider frame structure information while modeling temporal features.The experimental results on three widely used datasets demonstrate that the proposed two algorithms perform better than eight existing algorithms,and STA-Xcep-LSTM is more suitable for users who are sensitive to inference time.