Fake Face Detection And Identification Of Celebrities Based On Semantic Segmentation

With the gradual development of fake face manipulation technique,the potential negative effects and threats brought by Deepfake products,especially for politicians,have become a serious topic in recent years.Therefore,Deepfake detection is crucial to the protection of personal privacy and security and the maintenance of social security.The current Deepfake detectors perform well,but are still limited by redundant information.These detectors contain deceptive information such as background areas and facial regions without texture,which occupies resources and affects detection accuracy.Therefore,this paper proposes two methods for guiding the detection network to focus on the face region to address the problem.The main contributions of this work are as follows:(1)A Deepfake detection method based on semantic segmentation of faces is proposed.For the target person,a manually labeled face semantic tag is used with the original image,and a mask image is automatically generated by a semantic segmentation network,which is later concatenate with the original image to generate a face semantic mask.Using this attention-guided data augmentation method can effectively help the classification network to extract features accurately and thus help detection.(2)A Deepfake detection method for face semantic region erasure based on classifier sensitive regions is proposed.Based on the face semantic label masks generated by method(1),the face semantic regions that are more sensitive to the classification network results are erased and the classification network is trained again,which enables to obtain a more robust classifier.(3)This thesis conducts sufficient experiments on both the dataset produced in this paper targeting national leaders and the public dataset in the domain.The two methods proposed in this thesis are compared with multiple Deepfake methods on different datasets.The experimental results show that both methods proposed in this paper achieve better detection performance in the Deepfake detection task and achieves superior performance to the state-of-the-art method on a subset of two different datasets.