TMFD: Two-Stage Meta-learning Feature Disentanglement Framework for DeepFake Detection

Xiaogang Zhu, Bo Lin, Xinan He, Jianfeng Xu, Feng Ding

Abstract: With the advancement of deep neural networks (DNN), detecting falsified facial images has become a significant challenge. Previous state-of-the-art methods used fake images generated by various techniques as positive samples during the training phase. This strategy ensures that the model is exposed to a wide range of fake generation methods, resulting in a substantially higher number of positive samples than negative ones. However, in real-world scenarios, authentic images distinctly outnumber fake ones, which is not consistent with the hypothesis of designing DNN-based forensics detectors. Furthermore, the excessive use of certain fake images for training DNNs may lead to overfitting to specific fake generation techniques. Current research indicates that detectors often learn semantic features of the images, which can affect their performance. To address this issue, we propose a two-stage meta-learning feature disentanglement framework (TMFD). By leveraging the prior knowledge of meta-learning and incorporating a fine-tuning stage, extensive experimental results demonstrate that our method performs excellently across different domains.