Cross-Domain Generalization Limits of Vision Foundation Models in Facial Deepfake Detection
Researchers have evaluated the effectiveness of vision foundation models in detecting facial deepfakes across different generative techniques. Their study compared three distinct learning paradigms: supervised macro-semantic features, self-supervised geometric features, and multi-teacher agglomerative representations. The findings indicate that while these models can identify entire face syntheses, they struggle with localized editing techniques when evaluated using linear probing. AI
IMPACT Highlights limitations in current AI models for detecting sophisticated deepfakes, indicating a need for more robust generalization capabilities.