Research

Paper

AI LLM March 10, 2026

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

Authors

Chao Shuai, Zhenguang Liu, Shaojing Fan, Bin Gong, Weichen Lian, Xiuli Bi, Zhongjie Ba, Kui Ren

Abstract

AI-generated image detection has become increasingly important with the rapid advancement of generative AI. However, detectors built on Vision Foundation Models (VFMs, \emph{e.g.}, CLIP) often struggle to generalize to images created using unseen generation pipelines. We identify, for the first time, a key failure mechanism, termed \emph{semantic fallback}, where VFM-based detectors rely on dominant pre-trained semantic priors (such as identity) rather than forgery-specific traces under distribution shifts. To address this issue, we propose \textbf{Geometric Semantic Decoupling (GSD)}, a parameter-free module that explicitly removes semantic components from learned representations by leveraging a frozen VFM as a semantic guide with a trainable VFM as an artifact detector. GSD estimates semantic directions from batch-wise statistics and projects them out via a geometric constraint, forcing the artifact detector to rely on semantic-invariant forensic evidence. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art approaches, achieving 94.4\% video-level AUC (+\textbf{1.2\%}) in cross-dataset evaluation, improving robustness to unseen manipulations (+\textbf{3.0\%} on DF40), and generalizing beyond faces to the detection of synthetic images of general scenes, including UniversalFakeDetect (+\textbf{0.9\%}) and GenImage (+\textbf{1.7\%}).

Metadata

arXiv ID: 2603.09242
Provider: ARXIV
Primary Category: cs.CV
Published: 2026-03-10
Fetched: 2026-03-11 06:02

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2603.09242v1</id>\n    <title>When Detectors Forget Forensics: Blocking Semantic Shortcuts for Generalizable AI-Generated Image Detection</title>\n    <updated>2026-03-10T06:16:35Z</updated>\n    <link href='https://arxiv.org/abs/2603.09242v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2603.09242v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>AI-generated image detection has become increasingly important with the rapid advancement of generative AI. However, detectors built on Vision Foundation Models (VFMs, \\emph{e.g.}, CLIP) often struggle to generalize to images created using unseen generation pipelines. We identify, for the first time, a key failure mechanism, termed \\emph{semantic fallback}, where VFM-based detectors rely on dominant pre-trained semantic priors (such as identity) rather than forgery-specific traces under distribution shifts. To address this issue, we propose \\textbf{Geometric Semantic Decoupling (GSD)}, a parameter-free module that explicitly removes semantic components from learned representations by leveraging a frozen VFM as a semantic guide with a trainable VFM as an artifact detector. GSD estimates semantic directions from batch-wise statistics and projects them out via a geometric constraint, forcing the artifact detector to rely on semantic-invariant forensic evidence. Extensive experiments demonstrate that our method consistently outperforms state-of-the-art approaches, achieving 94.4\\% video-level AUC (+\\textbf{1.2\\%}) in cross-dataset evaluation, improving robustness to unseen manipulations (+\\textbf{3.0\\%} on DF40), and generalizing beyond faces to the detection of synthetic images of general scenes, including UniversalFakeDetect (+\\textbf{0.9\\%}) and GenImage (+\\textbf{1.7\\%}).</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.CV'/>\n    <published>2026-03-10T06:16:35Z</published>\n    <arxiv:primary_category term='cs.CV'/>\n    <author>\n      <name>Chao Shuai</name>\n    </author>\n    <author>\n      <name>Zhenguang Liu</name>\n    </author>\n    <author>\n      <name>Shaojing Fan</name>\n    </author>\n    <author>\n      <name>Bin Gong</name>\n    </author>\n    <author>\n      <name>Weichen Lian</name>\n    </author>\n    <author>\n      <name>Xiuli Bi</name>\n    </author>\n    <author>\n      <name>Zhongjie Ba</name>\n    </author>\n    <author>\n      <name>Kui Ren</name>\n    </author>\n  </entry>"
}