Suspicious Frequency Amplified Hybrid Framework for Generalizable AIGC Image Detection

The rapid advancement of generative adversarial networks and diffusion-based models has enabled the creation of highly realistic synthetic images, raising urgent concerns regarding the authenticity of multimedia data. While existing detection methods rely predominantly on spatial or frequency features, their performance often degrades under distribution shifts or novel generative models. In this work, we propose a comprehensive AIGI detection framework that integrates three complementary modules to exploit diverse artifact representations. The Noise Fingerprint Extraction and Enhance (NF2E) module captures sensor-level inconsistencies in residual noise patterns, providing a distinguishable low-level fingerprint for distinguishing real from synthetic images. The Suspicious Frequency Attention Module (SFAM) identifies abnormal spectral distributions by adaptively emphasizing informative frequency bands, enabling the detector to exploit subtle, non-trivial discrepancies across generative models. Complementing these, the Gaussian Mixture Patch Selection (GMPS) module introduces learnable Gaussian priors into patch-level attention in the CLIP visual branch, selectively highlighting suspicious spatial regions and enhancing feature separability. Extensive experiments across GAN- and diffusion-based datasets demonstrate that our method not only achieves strong generalization against unseen generative models. The codes will be released on https://github.com/wangjun9276/AIGI_SFAHDetection