Title: Enhancing Generalization in Deepfake Detection by Mitigating Forgery-Specific Shortcuts

Abstract

Current deepfake detection systems often struggle with generalization across different forgery techniques, primarily because existing models depend on spurious, method-specific shortcuts that do not transfer effectively to unseen manipulations. Although recent efforts have aimed to boost generalization capabilities, they have yet to incorporate an explicit mechanism for identifying and suppressing these shortcuts within learned representations. To address this gap, we introduce the Shortcut Subspace Suppression (S³) framework, which utilizes subspace modeling to explicitly characterize and mitigate method-specific shortcuts.

Our central hypothesis is that the variations distinguishing various forgery methods capture specific artifacts, serving as a reliable proxy for method-specific shortcuts. To operationalize this, we employ a lightweight linear probe for forgery method classification and apply Singular Value Decomposition (SVD) to isolate the dominant shortcut subspace. Leveraging this formulation, we implement two complementary strategies to diminish reliance on these shortcuts.

During the training phase, we apply soft suppression to the shortcut subspace within feature representations. This encourages the model to prioritize more generalizable cues for distinguishing real from fake content. At inference time, we offer a training-free alternative that attenuates neurons aligned with the identified shortcut directions. This approach allows for plug-and-play enhancement of generalization while improving model interpretability. Comprehensive experiments across multiple benchmarks confirm that our method significantly boosts cross-method generalization without compromising strong in-domain performance. The source code will be made available upon acceptance of the paper.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC