The Research Of Face Forgery Detection Based On Global Interaction And Local Discrepancy

Nowadays,the rapid development of deep learning has greatly contributed to the advancement of face forgery technology,not only lowering the threshold of using forgery tools,but also improving the quality of face generation images.Especially after the generation adversarial network has been proposed,various generation technologies have emerged and gradually become more and more perfect.The fake digital images are becoming more and more realistic.And the operation of face replacement with any software will no longer require professional video image producers,but this will pose a great threat to personal privacy,social ethics and even national security.Therefore,it is particularly important to detect the fake media data generated by face forgery technology.To address the shortcomings of existing face forgery detection techniques,two areas of research work were completed:(1)To address the problem that existing face forgery detection techniques do not consider image content interaction and compression processing to lose forgery traces,this paper proposes a deep face forgery detection method based on high-frequency features and global interaction.The spatial interaction between pixels is captured by scaling local self-attention mechanism to achieve global interaction.In addition,high-pass filters are used to amplify high-frequency subtle artifacts to improve the detection accuracy of compressed forged content.It also combines spatial domain and frequency domain information to maximize the capture of image semantic information effectively improve the generalization of the model.(2)For the face forgery detection methods with local forgery traces ignore the different effects of different image patchs on face operation detection,this paper proposes a face forgery detection method based on local disparity feature learning.Using the patch disparity feature module,the disparity features of genuine and fake faces are extracted by learning the feature representation of the patch considering the importance of the patch.The texture enhancement module is also used to prevent the loss of shallow texture features,which enables the model to extract more discriminative features.Finally,the two features are concated and discriminated by a classifier.The above two methods are experimentally validated on two large-scale fake face datasets,Face Forensics++ and Celeb-DF,and demonstrate the effectiveness of the different modules through ablation experiments.The experimental results show that the proposed method in this paper exhibits better detection results than the previous methods,and also achieves higher accuracy in cross-dataset detection.Figure 26 Table 11 Reference 72...