ForgeReason: Multimodal Reasoning with Multi-Agent Collaboration for Generalizable and Explainable Image Forgery Detection

Abstract The rapid advancement of generative AI and sophisticated image editing tools has made image forgery increasingly convincing, posing serious threats to media integrity, legal evidence, and public trust. Existing forgery detection methods primarily rely on single-modality analysis and operate as black-box classifiers, suffering from limited generalization to unseen forgery types and a lack of interpretability in their detection decisions. In this paper, we propose ForgeReason, a multimodal reasoning framework that integrates dual-stream feature extraction, reasoning alignment, and multi-agent collaboration for generalizable and explainable image forgery detection. The dual-stream feature extractor simultaneously captures spatial semantic features through a vision transformer and frequency-domain forensic traces through SRM-filtered encoding, with multi-scale attention fusion to enhance manipulation-sensitive representations. The reasoning alignment module bridges visual forensic features and linguistic reasoning spaces via cross-modal contrastive learning, enabling the model to articulate detection evidence in natural language. The multi-agent collaborative decision mechanism coordinates specialized agents for visual analysis, frequency analysis, and reasoning verification, producing comprehensive and reliable detection outcomes with interpretable explanations. Extensive experiments on three benchmark datasets demonstrate that ForgeReason achieves state-of-the-art performance, obtaining 96.8 percent accuracy on CASIA v2, 90.3 percent on IMD2020, and 92.5 percent on AutoSplice, while providing high-quality textual explanations for its detection decisions. Ablation studies confirm that each component contributes positively, with the full framework outperforming all individual and partial configurations.