Strong Generalization Deep Fake Face Detection Algorith

The wave of artificial intelligence has promoted the development of image generation and editing technology based on deep learning,which is the origin of deep face forgery technology.This technology mainly realizes the forgery of face images and videos through deep learning algorithm.The spread of forged faces violates the authenticity and uniqueness of face as a natural identity attribute,and poses a serious challenge to personal property,social stability and even national defense security.The existing detection work mainly detects the specific forged traces in the forged face by learning and lacks the mining of the essential features of real and fake faces,resulting in certain limitations in the generalization of the algorithm.Therefore,aiming at the problem of generalization,this paper makes a deeper exploration.This paper mainly studies strong generalization face forgery detection algorithms,and proposes three specific solutions:(1)Aiming at the problem of low generalization of the small deep forgery detection models,a strong generalization detection algorithm based on attention local contrast learning is proposed,which projects the local features of a face image onto a unit hypersphere,and enhances the forgery identification ability of the detection model by introducing a contrast learning mechanism;This scheme has good generalization and robustness for compressed videos,and the proposed algorithm does not rely on large-scale deep neural networks,and only uses small networks such as Res Net18 to achieve the same effect.(2)Aiming at the problem of weak generalization of face features,a strong generalization detection algorithm based on local anomalies is proposed.By learning the second-order similarity relationship between local regions in the depth feature map,the model is forced to learn the essential source feature differences of real and fake images,improving the accuracy of authenticity identification.In terms of intra domain testing,the method in this paper achieves over 99.5% AUC in multiple data sets with image scales of over 100000 orders of magnitude,such as Face Forensics++,Celeb-DF,DFD,and DFDC.In terms of cross domain testing,the average AUC of mutual cross domain testing in the four internal subsets of Face Forensics++,Deep Fake,Face2 Face,Face Swap,and Neural Textures,exceeds 95%.(3)Aiming at the problem that the deep forgery detection model is easy to overfit,a strong generalization detection algorithm based on mask learning is proposed,and a method that does not rely on data augmentation is proposed to extend the generalization boundaries learned by the model.Through the teacher student architecture,the easy over fitting features learned by the teacher model are shielded,reducing the risk of over fitting,and improving the ability of generalization detection.This scheme has good results in compression domain generalization,can effectively resist heavy video compression,and does not rely on any form of external data synthesis training.