Taming the Forensic Singularity: A Regularized Hyperbolic Framework for Generalizable AI-Generated Image Detection

ICLR 2026 Conference Submission107 Authors

Keywords: Multimedia Forensics, AI Generated Image Detection

Abstract: Detecting AI-generated images is a critical task in multimedia forensics, yet the generalization of detectors to unseen generative models remains a persistent challenge. While pre-trained Vision Transformers (ViTs) have emerged as powerful feature extractors, existing forensic methods often default to using final-layer features, which may discard crucial forensic traces. We embark on a systematic probe into the latent representations of various ViTs and uncover a universal phenomenon we term the "Forensic Singularity": a narrow region within the ViTs' mid-level layers where forensic separability culminates before giving way to semantic abstraction. To harness the immense potential of this "singularity" layer while mitigating its high risk of overfitting to limited training data, we propose a Regularized Hyperbolic Framework. Our framework learns a polar representation in the Poincaré Ball by disentangling semantic content and forensic evidence: the final-layer semantic feature guides the direction, while the singularity-layer forensic feature determines the radius. This inherent geometric constraint regularizes the model, promoting a more generalizable decision boundary. Extensive experiments demonstrate that our approach not only establishes a new state of the art on multiple datasets, and exhibits superior generalization performance on unseen generative models. Our work provides both a powerful new tool for AI forensics and a deeper insight into how the hierarchical representations of ViTs can be effectively harnessed.

Primary Area: applications to computer vision, audio, language, and other modalities

Submission Number: 107