Bi-level graph reasoning with frequency guidance and spatial constraint for AIGC-manipulation detection and localization

With the rapid advancement of generative models in image synthesis and editing, manipulated content increasingly resembles real images in both visual quality and statistical distribution. This presents new challenges for multimedia forensics. Traditional detection methods that rely on low-level statistical anomalies or local semantic inconsistencies are becoming less effective in complex generative manipulation scenarios. Achieving accurate forgery localization under AI-generated editing scenarios has become an urgent problem. To address this challenge, we propose a method termed Bi-level Graph Reasoning with Frequency Indication and Spatial Constraint (BiGR-Net) for Artificial Intelligence Generated Content (AIGC) manipulation detection and localization. By combining frequency-guided cues with spatial constraints, the proposed approach captures both the discriminative features of manipulated regions and their overall consistency. Specifically, through bi-level graph relation modeling driven by frequency guidance and spatial constraints, the method enhances forensic cues from both local anomaly and global consistency perspectives, enabling pixel-level localization and image-level manipulation detection. Extensive experiments on the AutoSplice and CelebA-HQ benchmark datasets show that the proposed method outperforms existing approaches in terms of localization accuracy, classification performance, and robustness. These results demonstrate the effectiveness of our approach in AIGC-manipulation scenarios.