Structural Semantic Features for Improved AI-Generated Fake Image Detection

ICLR 2026 Conference Submission22643 Authors

Keywords: AI-generated image detection, Structural semantics, Hierarchical image partitioning, Computer Vision

Abstract: The proliferation of AI-generated content (AIGC) has made the accurate detection of fake images a critical challenge. Existing state-of-the-art methods, such as PatchCraft and AIDE, primarily leverage local features like patch-wise frequency information or global semantic features derived from large-scale models like CLIP. While effective, these approaches often fail to incorporate the underlying structural semantics of an image, which are crucial for detecting the subtle inconsistencies and artifacts left by generative models. We propose a novel approach that augments existing AIGC detection frameworks by explicitly incorporating structural semantic information. Our method employs cuboidal partitioning, a hierarchical tool that recursively divides an image into meaningful sub-regions. At each division, we extract a measure of the statistical difference between the parent and child segments, which are then integrated with AIDE's existing features. Experimental results demonstrate our model's superior performance. We establish a new state-of-the-art in mean accuracy on the GenImage benchmark, proving our effectiveness on modern diffusion models. Our method also shows strong generalization by achieving second-best overall mean accuracy on the diverse AIGCDetect benchmark and a second-place finish on the challenging Chameleon dataset. These results highlight the significant value of structural semantics for building robust and generalizable AIGC detectors.

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

Submission Number: 22643