Omniaggregation Topology Perception Networks for AIGC-Manipulation Detection and Localization

As artificial intelligence generated content (AIGC) technologies become increasingly accessible, the risk of malicious image manipulation increases, which poses new challenges to the reliability of industrial visual systems. Compared to traditional tampering, AIGC-edited images often disrupt the semantic topology and structural consistency of visual content, making forgery localization more difficult. To address this, omniaggregation topology perception networks (OTP-Net) is proposed, a novel framework designed to jointly capture pixel-level artifacts and topological inconsistencies. Specifically, a high-frequency discrepancy learning module is introduced to enhance fine-grained manipulation traces, while spatial and channel topology perception branches are designed to capture multiscale topological anomalies through graph-based modeling. Extensive experiments demonstrate that OTP-Net achieves competitive performance in CelebA-HQ and AutoSplice datasets, while showing notable gains in the average pixel-level F1 and intersection over union by up to 3.86% and 4.42% in challenging composite attacks.