Generalizable Deepfake Detection with Unbiased Feature Extraction and Low-Level Forgery Enhancement

Zhihan Yu, JiaXin Li, Guangshuo Wang, Yuesheng Zhu, Guibo Luo

Abstract: Deepfake detection has recently become an urgent issue since the deepfake technology has raised security concerns in society. However, current deepfake detection methods exist susceptibility when encountering unseen data, limiting their generalization ability. In this paper, we propose a straightforward yet effective framework of deepfake detection based on unbiased feature extraction and low-level forgery enhancement (UFELE). Obtaining unbiased features utilizing frozen Visual Foundation Models, we devise a Low-level Forgery Feature Enhancement (LFFE) module to extract and enhance the low-level features from the frozen intermediate block. Also, an Adaptive Feature Fusion (AFF) module is designed to amalgamate the enhanced low-level forgery features with the high-level semantic features flexibly. Extensive experiments on several datasets illustrate that our proposed method has better detection performance than the state-of-the-art methods in terms of generalizability and robustness.