Toward Calibrated, Fair, and accurate Deepfake Detection

In three linesFace-Fairness (FF) is a plug-and-play framework to reduce demographic bias in deepfake detection. Face-Feature Tuning (FFT), its main contribution, remaps logits without demographic labels using frozen embeddings. FF-Max and FF-Discover optimize minority group accuracy. Tests show reduced FPR/TPR gaps and improved overall accuracy across in-domain and cross-dataset settings.

Summary generated by Claude — human-verified