Train only once: FADNet detects face authenticity with uncertainty-aware one-class learning

FADNet reframes face forgery detection as a one-class classification task, training only once on real face data; features that deviate significantly from the real distribution are judged as forged, with no need to retrain for each forgery method. The framework uses evidential deep learning (EDL) to quantify prediction uncertainty and a plug-and-play pseudo-forged image generator (PFIG) to tighten the decision boundary. On the DF40 and ASFD benchmarks, FADNet achieves an average accuracy of 96.63% and an average precision of 98.83%, outperforming existing SOTA methods.