Knowledge Enhanced Dynamic Trust Network for Multimodal Fake News Detection

With the rapid rise of social media, fake news spreads swiftly through text–image combinations, threatening the information ecosystem and social stability. Multimodal fake news detection has thus become a research hotspot. Existing methods that fuse textual, visual and propagation signals face two key challenges: Semantic inconsistencies across modalities lead to biased results and uneven reliability within modalities, where fixed-weight fusion is vulnerable to low-quality or manipulated inputs. To address these issues, this paper proposes a Knowledge-Enhanced Dynamic Trust Network (KDTN). Its contributions are twofold. First, a knowledge-enhanced tri-consistency mechanism evaluates global consistency among text–image, text–knowledge and image–knowledge pairs, while an inverse attention mechanism captures local contradictions. This design strengthens detection from both global and local perspectives. Second, a dynamic modality trust fusion mechanism accounts for uncertainty, cross-modal divergence and source priors. Through a gated network, it allocates trust dynamically at the sample level, reducing the impact of unreliable modalities. KDTN thus maintains overall consistency constraints, highlights conflicts and adaptively adjusts modality importance, improving robustness and interpretability. Experiments show KDTN outperforms mainstream baselines on multiple datasets, boosting accuracy and robustness and effectively localizing contradictory regions in explainability analysis.