TERN: Type-Aware Evidence Reasoning for Multimodal Fake News Detection

The proliferation of multimodal fake news on social media poses a persistent threat to public opinion and information credibility. Existing multimodal detection methods typically adopt a unified fusion strategy for all samples and learn entangled representations where deceptive patterns and semantic content are mixed. As a result, they struggle to (i) account for heterogeneous deception mechanisms such as image tampering and text--image mismatch, and (ii) perform hierarchical reasoning over multiple sources of evidence. To address these limitations, we propose TERN, a type-aware evidence reasoning network for multimodal fake news detection. TERN is built upon three key components. First, a type discovery module leverages contrastive learning and prototype-based clustering in the image feature space to automatically uncover latent deception types without manual annotations. Second, a type--semantic disentanglement module explicitly separates type-discriminative information from content semantics, mitigating spurious correlations between topics and veracity. Third, a type-guided hierarchical evidence fusion module generates textual, visual authenticity, and cross-modal consistency evidence, and adaptively integrates them via type-conditioned attention. Experiments on four public benchmarks---MR2-Chinese, MR2-English, Weibo, and PHEME---demonstrate that TERN outperforms strong multimodal baselines. On average, TERN achieves 93.21\% accuracy and 91.09\% F1-score, with clear gains in Matthews correlation coefficient, indicating more balanced decisions under class imbalance.