Generalizable Deepfake Detection via Frequency-Domain Enhancement and Feature Disentanglement

Abstract

Existing deepfake detectors often perform well on in-domain data but generalize poorly to unseen datasets or manipulation methods. This limitation is largely attributed to their reliance on dataset-specific semantic cues rather than transferable forgery patterns. To address this limitation, we propose a generalizable deepfake detection framework that combines frequency-domain enhancement with feature disentanglement. A Phase-Amplitude Frequency Enhancement (PAFE) module enhances subtle spectral artifacts introduced during deepfake generation. We then feed the enhanced representations into an asymmetric dual-branch architecture that separates content-related information from forgery-related features. The content branch models facial semantics, while the forgery branch extracts discriminative forgery features with reduced content interference. A spatial self-attention module further refines the forgery features. We optimize the framework using image-level reconstruction loss, feature-level contrastive loss, and classification loss. Together, these objectives encourage effective feature disentanglement and improve the discriminability of the learned forgery features. Extensive experiments on several widely used deepfake benchmarks show that the proposed framework achieves competitive detection performance and improved cross-domain generalization compared with existing methods.