Title: When Detectors Forget Forensics: Blocking Semantic Shortcuts for Generalizable AI-Generated Image Detection

Abstract: As generative models produce increasingly realistic outputs, distinguishing between authentic and synthetic imagery has become increasingly difficult, presenting major hurdles for accurate AI-generated image detection. While large-scale pre-trained Vision Foundation Models have significantly improved detection performance, their ability to generalize to images created by unseen generation pipelines remains insufficient. This study identifies, for the first time, a critical failure mode called semantic fallback, where forensic fine-tuning does not completely restructure the representation space. As a result, the learned representations continue to cluster around high-level semantic features rather than focusing on specific forensic indicators of manipulation. Guided by this discovery, we introduce the Geometric Semantic Decoupling (GSD) framework, designed to explicitly dampen semantically dominant directions to foster invariant forensic representations. GSD employs a frozen CLIP encoder to identify the primary semantic subspace using Singular Value Decomposition (SVD). It subsequently reduces these semantic components through a geometry-constrained approach, with suppression intensity dynamically adjusted across different samples and network layers. To address efficiency, we present a mini-batch SVD approximation method that spreads the cost of subspace estimation, resulting in a computational overhead reduction of more than $15 \times$ without compromising performance. Lastly, to accommodate practical applications in both large-scale and online evaluation settings, we establish three inference protocols: batch, per-sample, and reference-based inference. We show that these methods achieve consistent semantic decoupling, leading to a stable feature manifold oriented toward forgery detection.

Source: arXiv Generated at: 2026-06-04 00:00:00 UTC