MAGNet: A multimodal knowledge-augmented graph network for early-stage misinformation detection

With the rapid proliferation of multimodal misinformation on social media, detecting such content has become increasingly challenging. Existing approaches often rely on flat or shallow fusion strategies, which fail to capture structured semantic interactions across modalities. Moreover, most methods lack controllable, task-relevant mechanisms for integrating external knowledge, limiting their adaptability to emerging misinformation. In this paper, we present MAGNet, a Multimodal Augmented Graph Network that models fine-grained features with LLM-enhanced contextual knowledge through a hierarchical graph attention framework. MAGNet constructs heterogeneous graphs with modality- and context-specific edge weights based on semantic and affective alignment, enabling progressive reasoning from local features to global representations. Extensive experiments on three real-world datasets demonstrate that MAGNet consistently outperforms strong baselines across multiple evaluation metrics. The results underscore the effectiveness of combining graph-based modeling, fine-grained fusion, and structured knowledge integration in developing scalable and robust solutions for multimodal misinformation detection.