Motion Magnified 3-D Residual-in-Dense Network for DeepFake Detection

Driven by the advances in deep learning, highly photo-realistic techniques capable of switching the identity and expression of faces have emerged. Cheap access to computing has brought such technology within the reach of anyone with a computer and Internet including people with sinister motives. To detect these forgeries, we present a novel compression resilient approach for deepfake detection in videos. The proposed approach employs motion magnification as a pre-processing step to amplify temporal inconsistencies common in forged videos. Utilizing these processed videos, we propose the 3D Residual-in-Dense ConvNet, which captures low level spatiotemporal features, which help classify a video as pristine or forged. The proposed method yields more than 93% average detection accuracy on the high compression variant of the FaceForensics++ dataset and achieves state-of-the-art performance on multiple benchmarks across the FaceForensics++ and CelebDF datasets. Further, we study the behavior of deepfake detection algorithms across ethnicities and demonstrate how the proposed method reduces the inherent bias against minority ethnicities prevalent in existing algorithms.