Mitigating Hierarchical Shortcut Reliance in Frozen CLIP for Cross-Generator AI-Generated Image Detection

<title>Abstract</title> <p>Detectors of artificial intelligence (AI)-generated images often perform well on familiar generators but degrade on unseen generators. This paper studies cross-generator generalization from the perspective of how a frozen vision foundation model uses representations across layers. Layer-wise probing shows that different generators favor different layers, while frequency-sensitivity analysis reveals different layer--frequency responses. These observations motivate the term hierarchical shortcut reliance: a detector may depend excessively on a small set of discriminative but weakly transferable layers and underuse complementary evidence from other layers. We propose a layer-wise debiasing framework based on a frozen Contrastive Language--Image Pre-training (CLIP) vision transformer (ViT). The framework applies spatial consistency intervention to low layers, frequency-prior intervention to middle layers, and semantic debiasing to high layers. Training additionally uses token-shuffle and token-mix branches to reduce position and content shortcuts, together with forgery-only latent augmentation. Experiments on UniversalFakeDetect and GenImage, robustness tests under common post-processing operations, and ablation studies show that the framework provides competitive cross-generator detection and that its gains are consistent with complementary layer-wise interventions. The conclusions are limited to the evaluated benchmarks and frozen CLIP ViT-L/14 setting.</p>