MADL: Towards Dependable Image Forgery Detection Via Multi-Agent Forensic Reasoning

The rapid advancement of generative models in creating realistic synthetic and manipulated images poses growing risks to visual-content authenticity. A practical forensic system must determine whether an image is authentic or forged, distinguish fully synthetic content from local manipulation, and locate tampered regions. Existing approaches typically rely on a single model for classification or localization, making it difficult to combine high-level semantic understanding with low-level manipulation traces when different evidence sources conflict. To address these challenges, we proposeMADL, a multi-agent collaborative framework for image forgery detection and localization. MADL extends single-model discrimination into a hierarchical, evidence-driven collaborative adjudication process. It comprises three complementary agents: Agent A performs Qwen-based semantic verification and image-level three-class classification; Agent B extracts pixel-level manipulation evidence through dual-stream forensic analysis and candidate-mask ranking; and Agent C coordinates heterogeneous evidence through conflict-aware adjudication. Through structured evidence interaction and explicit agreement, suppression, override, and conflict-handling mechanisms, MADL combines semantic reasoning with manipulation-trace localization while keeping the decision process traceable. Experiments demonstrate competitive three-class classification and strong IoU-based tampered-region localization. Further ablations and decision-path audits validate the multi-agent collaboration and conflict-aware adjudication mechanisms.