Improving Generalization in Deepfake Detection via Augmentation with Recurrent Adversarial Attacks

The crucial effort to counteract deepfakes and misinformation at large holds great importance in our society, especially at this moment in time. Deepfake detectors evolve at the same pace as deepfake generators, or even slower, and more than that, they are trained on a limited amount of data and do not achieve generalization in most situations. The primary challenge associated with training deepfakes lies in the necessity for a substantial number of diverse generated samples originating from a multitude of distinct models — an achievement not easily attained. For that reason, this work aims to leverage the one abundant resource at our disposal: real videos. This paper presents a novel training framework for deepfake detectors, aimed to improve generalization by continuously using adversarial attacks to generate new deepfakes that a detector might not be trained to recognize, starting from real samples. We aim to do that while keeping the generated samples as realistic as possible. We train the deepfake detector on the newly generated deepfakes, along with the original images, aiming to enhance its ability to differentiate between them. We show that this training method improved generalization to unseen datasets, while not using new data. More than that, this unsupervised method only uses real images, making it an easy to implement and adaptable way to improve generalization.