MH-FFNet: Leveraging mid-high frequency information for robust fine-grained face forgery detection

The rapid advancement of Deepfake technology has rendered the generation of forged faces highly realistic, while simultaneously introducing significant societal security concerns. The accurate detection of forged facial images has thus emerged as an urgent issue and a formidable challenge. In this paper, we approach face forgery detection as a fine-grained classification problem due to the subtle differences between real and fake faces. We propose a detection framework termed the Mid-High Frequency Based Fine-Grained Network (MH-FFNet), which enhances the detection of forged faces by leveraging mid- and high-frequency information to capture fine-grained forgery cues. To better extract and utilize these cues, we devise two fine-grained feature enhancement modules: the Patch-based Fine-Grained Enhancement Module (P-FGEM) and the Feature-based Fine-Grained Enhancement Module (F-FGEM). The P-FGEM module focuses on extracting mid- and high-frequency information from shallow feature blocks, enhancing forgery representations in shallow features. This design effectively mitigates the loss of mid- and high-frequency cues as the network deepens, thereby improving the algorithm’s sensitivity to forgery cues. In contrast, the F-FGEM module captures mid- and high-frequency information from mid-level global features, further enriching forgery representations in these features and significantly enhancing their discriminative power. Experimental results indicate that our proposed method achieves an AUC of 99.44% on the FF++ (C23) dataset and 83.44% on the Celeb-DF (V2) dataset, demonstrating the algorithm’s superior detection capability and generalization performance. Additionally, we conduct experiments to comprehensively illustrate the robustness of the algorithm against common image post-processing attacks.