VLAForge: Unleashing vision-language semantics to detect deepfake videos

The VLAForge framework uses cross-modal semantics in pretrained vision-language models such as CLIP to detect deepfake videos, significantly outperforming existing state-of-the-art methods in both frame-level and video-level evaluations. Its ForgePerceiver module captures subtle forgery cues while preserving pretrained vision-language alignment knowledge, and an Identity-Aware VLA score enhanced by identity-prior text prompts serves as a supplementary discriminative cue. Experiments cover two types of forgery: traditional face swapping and recent full-face generation.