Learning a deep dual-level network for robust DeepFake detection

Highlights：

• We alleviate the adverse impact of an imbalanced dataset by introducing the AUC loss to directly maximize the AUC score and minimize the adverse impact of an imbalanced dataset.

• We combine the features of both frame-level and video-level methods into an integrated network that detects not only each individual forged frame but also takes the temporal relations of frames into consideration to provide an overall prediction for a whole video.

• By sampling the Celeb-DF and DFDC datasets, we generate five subsets to simulate real-world imbalanced data conditions.

• We conduct an exhaustive analysis on the data imbalance problem to highlight the efficacy of our proposed joint loss function and the robustness of our model when facing data imbalance problems.