目标域增强驱动的无监督域自适应人脸伪造检测OA

The generalization ability of current face forgery detection methods remained insufficient for deployment in real-world scenarios.To address this issue,this paper proposed an unsupervised domain adaptation(UDA)framework that enhanced detection performance on target domains with unseen forgery patterns.Specifically,considering the significant domain discrepancies introduced by different forgery generation methods,designed a cross-domain feature perturbation strategy that incorporated target domain features into the source domain,enabling indirect domain alignment.The adversarial learning module was applied to further align forged features from both domains.Additionally,introduced a discriminative clustering loss,which combined an entropy loss and a class-balancing loss.This loss simultaneously improved the model's transferability and ensured balanced classification.Extensive experiments on the DF,F2F,and FS datasets demonstrated that their approach consistently outperformed existing domain generalization and domain adaptation methods,achieving over 95%in both AUC and F1-scores on the target domains.