文献类型：期刊文献

英文题名：S2F-Net: A robust spatial-spectrum fusion framework for cross-model AIGC detection

作者：Hu, Xiangyu[1];Zheng, Hongchuang[2];Hong, Yicheng[1];Zeng, Wenjun[1];Liu, Bingyao[1]

机构：[1]Guangdong Ocean Univ, Coll Comp Sci & Engn, Yangjiang 529500, Peoples R China;[2]Guangzhou Univ, Sch Artificial Intelligence, Guangzhou 510623, Peoples R China

年份：2027

卷号：334

外文期刊名：EXPERT SYSTEMS WITH APPLICATIONS

收录：SCI-EXPANDED(收录号：WOS:001887095400001)、、WOS

语种：英文

外文关键词：AI-generated image detection; Spatial-spectrum fusion; Cross-model generalization; Frequency-domain artifacts; Learnable Frequency attention

外文摘要：The rapid advancement of generative models necessitates generalizable detection frameworks capable of identifying content from unseen architectures. However, existing detectors frequently overfit to domain-specific semantic distributions, causing substantial generalization decay. To overcome this bottleneck, we propose S2F-Net, a lightweight and generalizable spatial-spectrum fusion framework for cross-model AI-generated image detection. Grounded in the observation that cumulative upsampling operators induce differentiated spectral artifacts across regions of varying texture complexity, we introduce a Learnable Frequency Attention (LFA) mechanism to establish deep synergy between spatial textures and frequency representations. Unlike conventional methods relying on handcrafted or fixed filters, LFA incorporates a distance-initialized weight prior, dynamically recalibrating frequency responses across entropy-stratified regions to isolate subtle synthesis traces. This dual-domain collaboration enables S2F-Net to effectively decouple content semantics from intrinsic forgery fingerprints. Extensive evaluations on the AIGCDetectBenchmark across 17 mainstream generative models demonstrate that S2F-Net achieves a state-of-the-art detection accuracy of 90.49%, while operating with only 1.20M parameters, significantly outperforming competitive baselines in both generalization and efficiency.