Suppressing Forgery-Specific Shortcuts for Generalizable Deepfake Detection

Researchers have proposed a new framework, Shortcut Subspace Suppression (S^3), to improve the generalizability of deepfake detection models across different forgery methods.

Deepfake detection models often struggle to generalize across various forgery methods due to their reliance on method-specific shortcuts, according to a paper submitted to arXiv on June 1, 2026 ^[1]. The S^3 framework addresses this issue by explicitly characterizing and suppressing these shortcuts via subspace modeling. The method involves training a lightweight linear probe for forgery method classification and performing Singular Value Decomposition (SVD) to extract the dominant shortcut subspace. Two complementary strategies are employed to reduce shortcut reliance: softly suppressing the shortcut subspace during training and attenuating neurons aligned with the identified shortcut directions at inference time. Another study on arXiv also explored techniques to improve deepfake detection, achieving an AUC of 0.905 on Celeb-DF v2 by selecting semantically meaningful patch tokens and using a linear probe to classify retained regions ^[2]. The S^3 framework significantly improves cross-method generalization while maintaining strong in-domain performance.