Investigating the Impact of Visual Attention Models in Face Forgery Detection

With the recent rise of realistic face manipulation methods, building robust face tampering detection methods has become more important than ever before. Visual attention has played an important role in highlighting discriminative regions within input which is important for making accurate predictions. This manuscript presents a comparative study of several recently proposed visual attention models for the problem of face forgery detection. Specifically, five visual attention models namely, coordinate, selective kernel, triplet, CoT, and shuffle attention have been tested by integrating with a baseline deep learning model. The modified visually attentive architectures are trained and tested on popular public benchmark dataset FaceForensics++. The experimental results achieved by different attention approaches are compared. Additionally, the computational costs involved in each type of attention have also been discussed specifying the accuracy and computation tradeoff. Experimental results prove that Triplet Attention performs best by achieving accuracy scores of 0.9543 and 0.7190 on DF and NT categories of the FF++ dataset. Triplet attention is also extremely lightweight with only 1200 trainable parameters compared to the other attention modules under study.