Robust deepfake detection via invariant feature learning

Abstract

We propose an effective invariant feature learning algorithm for Deepfake detection that is robust to unseen Deepfake methods. Recently, elaborate face-swap techniques using deep learning, so-called Deepfakes, have emerged with the advanced deep learning-based image generation technology. As a result, it has been challenging to distinguish between the original face image and the swapped one with naked eyes. Moreover, social threats have been increasing due to cases of exploiting Deepfakes. Existing Deepfake detection methods exploit specific abnormal signals in Deepfakes. However, the performances were degraded for unseen Deepfake methods. This failure is because the detection model learned specific method-dependent features that mostly depend on the training environments rather than invariant features across various Deepfake methods. To resolve the issue, we propose an effective invariant feature learning algorithm, called Adversarial Invariant Risk Minimization (AIRM), which exploits adversarial learning and invariant risk minimization to extract invariant features across different Deepfake environments. We demonstrate that the proposed algorithm enables models to be more robust in unseen Deepfake environments. We also analyzed quantitatively and qualitatively that these detection models effectively learned invariant features. Furthermore, we suggest a fair data sampling technique that allows the model to avoid being biased in specific environments in the training process, which makes it possible to effectively extract invariant features.