Deepfake detection via Feature Refinement and Enhancement Network

Abstract

The rapid advancement of deepfake technology poses significant threats to the integrity and privacy of biometric systems, such as facial recognition and voice authentication. To address this issue, there is an urgent need for advanced forensic detection methods that can reliably safeguard biometric data from manipulation and unauthorized access. However, current methods mainly focus on shallow feature extraction and neglect feature refinement and enhancement, which leads to low detection accuracy and poor generalization performance. To address this problem, we propose Feature Refinement and Enhancement Network (FRENet) for deepfake detection by leveraging progressive refinement and enhanced mixed feature learning. Specifically, a Low Rank Projected Self-Attention (LPSA) module is introduced for the refinement and enhancement of features. Also, a Patch-based Focused (PatchFocus) module is proposed to highlight local texture inconsistencies in key regions. In addition, we propose a Refine Fusion (RefFus) module that integrates the refined features and associated noise information to enhance feature separability. Experimental results across five benchmark datasets demonstrate that the proposed FRENet outperforms state-of-the-art methods in terms of both accuracy and generalization. The code is available at https://github.com/weichengsong-code/FRENet.