WildDFDet: Robust Unsupervised DeepFake Detection in the Wild

Summary

WildDFDet is an unsupervised DeepFake detection framework built for in-the-wild forensic data, where labels are missing or noisy. It uses a vision foundation model as a Pseudo-Label Generator, then trains with Density-Aware Weighted Contrastive Learning and Density-Guided Label Correction to refine noisy pseudo-labels. On the authors' in-the-wild benchmark it reaches an average AUC of 98.21%, outperforming existing methods by 27.68%–45.59%.

Topics

• Deepfakes and impersonation