Analyzing Deep Learning Models’ Generalization Ability Under Different Augmentations on Deepfake Datasets

Abstract

Deepfakes allow users to manipulate the identity of a person in a video or an image. Improvements on GAN-based techniques also generate more realistic and hard to detect fake faces. This threatens individuals and decreases trust in social media platforms. In this work, our goal is to report eight different models’ learning ability on, by far, the largest fake face dataset - DFDC. The models’ generalization ability was tested on the DFDC test set and Celeb-DF-v2 dataset. Effect of the various cut-out like augmentations to the learning was also reported.