Research On Deep Fake Detection Algorithm With High Generalizatio

Deepfake algorithm is a technology that utilizes the powerful fitting ability of deep learning to generate realistic forgery faces.In recent years,the malicious videos generated by such algorithms have been widely spread in the network,which poses a serious threat to personal privacy and social stability.Therefore,detecting such deepfake videos has become an urgent problem to be solved.The related research on deepfake detection has gained sufficient high accuracy on various data sets,while the generalization performance is still insufficient.Different forgery algorithms leave different traces,and the unknown data distribution increases the difficulty of detection.Most of the existing methods are aimed at analyzing and detecting specific traces and distortions generated by a specific forgery algorithm.However,such detection algorithms usually have a significant decline in accuracy when detecting forgery videos generated by other algorithms.Therefore,this paper intends to study the generalization of deepfake detection algorithms.Considering different scenarios of model generalization,this paper designs two specific schemes.The research contents of this paper are as follows:(1)For the multi-domain generalization problem,this paper proposed a deepfake detection scheme based on commonality learning strategy.Existing deep forgery schemes are varied in the forgery generation process,while inevitably leaving some similar traces in forgery videos(boundary anomalies,PRNU noise,biological signals,etc).This scheme attempts to extract common features from multiple kinds of forgery data,and thus achieve better generalization on the unknown dataset.Specifically,a two-stage training scheme is proposed.Firstly,a Specific Forgery Feature Extractor(SFFExtractor)is trained to learn the specific distribution of each forgery dataset.Then,with the assistance of SFFExtractors,the Common Forgery Feature Extractor(CFFExtractor)is trained to learn the commonality that existed in the specific forgery features,so as to enhance the detection ability on the unknown forgery datasets.The experimental results show that the commonality learning strategy helps to improve the generalization performance of the detection algorithm,and the detection results are better than many existing advanced research results.(2)For the single-domain generalization problem,this paper proposed a Deepfake detection scheme based on the adversarial style transfer algorithm.The first scheme extracts the commonality between various forgery features,and thus generalizes well on unknown types of forgeries.However,the data acquisition process is not easy in the real world.Therefore,this paper considers the most difficult situation in the deepfake detection generalization problem: How to generalize to a variety of unknown forgery data when only the real data is known.The scheme defines the attribute of unknown forgery data as the difference of forgery scheme and the difference of image style,thus proposing the Universal Forgery Generation(UFG)and Adversarial Style transfer algorithm(AST)to extend existing datasets.The experimental results show that our scheme can simulate diverse unknown forgery data,and the generalization performance is superior to many existing schemes.