Research On Detection Technology Of Face Forgery Method With FaceShifter

With the development of technology,the public can easily and quickly access a large number of digital media content through mobile phones,tablets and other devices.But in recent years,facial-tampering techniques based on deep learning have emerged,putting the authenticity of digital media content in jeopardy.Among them,the fake facial content made by FaceShifter based on the GAN(Generative Adversarial Network)has become a significant faceswapping method because of its high credibility and fidelity.Due to the malicious use of this method,it poses serious threat to personal privacy and social stability.In this thesis,two methods based on deep learning are proposed to detect fake faces made by FaceShifter.The main works of this thesis are as follows:1.Building FaceShifter image dataset,which includes Full-image scenario containing environmental information(such as background,clothing)and facial information,and Faceonly scenario containing only facial information.Each scenario contains 64000 face images,which 32000 are real faces and 32000 are FaceShifter fake faces.The Full-image scenario dataset is made by using Open CV software library,which intercepts 1000 videos of real face and 1000 FaceShifter videos of fake face according to frames.Each video intercepts 32 frames,and obtains facial images containing environment information and facial information.Faceonly scenario dataset is made by MTCNN algorithm,which recognizes and extracts facial images of Full-image scenario to get facial images that filter out environment information and only contain facial information.2.Aiming at FaceShifter fake facial images,a detection algorithm based on fine-tuning and Xception is proposed.Xception is a kind of deep convolution neural network,and finetuning refers to applying the knowledge learned from other tasks to different but related tasks.The pre-training model of Xception trained with Image Net dataset is used to extract the features of the input facial image.the results show that: The accuracy in the face-only scenario is 94.62%,which is 13.16% higher than that in the full-image scenario;Compared with the mainstream methods of detection,Two-Stream,Meso-4,Meso Inception-4,Capsule-Forensics,the accuracy is increased by 21.98%,13.37%,12.43%,9.11% respeactively.3.In order to further improve the performance of detection,a detection algorithm based on attention mechanism and EfficientNet B4 is proposed.EfficientNet B4 is a scalable convolution neural network,and the attention mechanism is a data processing method.A selfattention layer is added between the 4th 5x5 MBCon V6 module and the 5th 3x3 MBCon V6 module of EfficientNet B4.The attention map of size 28x28x56 replaces the original feature map.The experiment results show that: The accuracy in the face-only scenario is 97.73%,which is 11% higher than that in the full-image scenario;Compared without self-attention module,the accuracy is increased by 2.52%.