Learning a deep dual-level network for robust DeepFake detection

Face manipulation techniques, especially DeepFake techniques, are causing severe social concerns and security problems. When faced with skewed data distributions such as those found in the real world, existing DeepFake detection methods exhibit significantly degraded performance, especially the AUC score. In this paper, we focus on DeepFake detection in real-world situations. We propose a dual-level collaborative framework to detect frame-level and video-level forgeries simultaneously with a joint loss function to optimize both the AUC score and error rate at the same time. Our experiments indicate that the AUC loss boosts imbalanced learning performance and outperforms focal loss, a state-of-the-art loss function to address imbalanced data. In addition, our multitask structure enables mutual reinforcement of frame-level and video-level detection and achieves outstanding performance in imbalanced learning. Our proposed method is also more robust to video quality variations and shows better generalization ability in cross-dataset evaluations than existing DeepFake detection methods. Our implementation is available online at https://github.com/PWB97/Deepfake-detection.