Fig. 5From: Multi-attention-based approach for deepfake face and expression swap detection and localizationROC curves illustrating the classification performance on (HQ—C-23) and (LQ—C-40) compression qualities from the FF++ datasetBack to article page