Don't Just Classify, Compare: Evidence-Driven Multimodal Experts Framework for Fake News Detection
Abstract
The proliferation of multimodal fake news, which combines misleading text with suggestive imagery, presents a significant societal challenge, demanding robust detection methods. Existing approaches have evolved from simple feature fusion to sophisticated cross-modal attention networks to capture inter-modality relationships. However, they still face critical limitations: 1) a failure to consult external knowledge, leaving them vulnerable to evolving narratives, and 2) an inability to represent the heterogeneous evidence patterns of multimodal samples, leading to overconfident errors on ambiguous cases. To address these gaps, we propose the Evidence-Driven Multimodal Experts (EDME), a framework that emulates human fact-checking by integrating retrieval-based evidence with Dirichlet-profile decision-making. EDME first employs a Multimodal Evidence Retriever to fetch relevant real and fake news articles, providing crucial external context. This evidence is then leveraged by specialized Evidence-Enhanced Multimodal Experts, whose judgments are dynamically aggregated by an Evidence-Driven Expert Decision mechanism that composes evidence-conditioned categorical profiles. Extensive experiments on three benchmark datasets demonstrate that EDME significantly outperforms state-of-the-art baselines, improving accuracy by up to 1.2% on the Weibo dataset.