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

2026-03-10

AI Summary

This work addresses the limited generalization of current vision foundation model–based AI-generated image detectors when confronted with unseen generation methods, as they often rely on semantic priors and overlook genuine forensic traces. To tackle this, the authors propose Geometric-Semantic Disentanglement (GSD), which for the first time identifies and mitigates the “semantic fallback” problem. GSD leverages a frozen vision foundation model to estimate semantic directions and explicitly removes semantic components from the learnable detector’s representation via geometric projection, thereby forcing the model to focus on semantic-agnostic forensic cues. Requiring no additional parameters, GSD achieves a video-level AUC of 94.4% in cross-dataset evaluation, surpassing the state of the art by 1.2%, improves robustness against unknown manipulations by 3.0% on DF40, and significantly outperforms baselines in general image detection tasks.

Technology Category

• Computer Vision: Other Foundations of Computer Vision
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning
• Machine Learning: Deep Neural Architectures and Foundation Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding

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\%}).

Problem

• AI-generated image detection
• semantic fallback
• generalization
• Vision Foundation Models
• forensic evidence

Innovation

• semantic fallback
• Geometric Semantic Decoupling
• Vision Foundation Models
• generalizable detection
• forensic evidence