Generalizable Face Forgery Detection via Perturb-and-Deblur Stability

The growing realism and spread of facial deepfakes threaten individual privacy and information security.However, existing face forgery detectors exhibit limited generalization capabilities, struggling to detect forgeries created by unseen manipulation methods during training. This limitation restricts the effectiveness of detectors in real-world applications. This paper introduces a novel detection framework built upon the concept of perturb-and-deblur stability. The key idea is that real and forged faces behave differently under perturbation followed by deblurring: real faces remain stable, while forged faces exhibit unstable reconstructions. To systematically exploit this property, deblurring-based auxiliary domains are constructed, including deblurred images, residuals, and hierarchical features, which are further enriched by a multi-scale attention mechanism. A transformer-based domain relation encoder integrates these domains to capture discriminative stability patterns for generalized classification. Extensive experiments conducted on multiple public benchmarks demonstrate that the proposed approach consistently outperforms state-of-the-art detectors, achieving superior robustness and cross-dataset generalization. These results verify the effectiveness of perturb-and-deblur stability as a generalizable cue for face forgery detection.