Three novel, identity-agnostic, generator-agnostic deepfake detection methods
built on top of the FairTalking-Bench dataset.
CTA
Causal Time-Arrow asymmetry (A->V vs V->A predictive loss gap)
Yes
TimeSformer / ViT + Wav2Vec2
PSM
Phoneme-Synchrony Manifold (phoneme-conditioned lip prior)
Yes
Wav2Vec2-phoneme + MediaPipe lip
RADR
Reference-Anchored Diffusion… See the full description on the dataset page:
https://huggingface.co/datasets/huahua123313/fairtalking-second-work.