Large Language Model Guided Graph Capsule Network with Local-Enhanced Alignment for Fake News Detection

Fake news on social media increasingly employs advanced deceptive strategies such as partial tampering and semantic misleading, which create a camouflaged consistency as it exhibits superficial coherence while harboring deeper inconsistencies. Existing detection methods treat cross-modal consistency as a monolithic concept and apply uniform strategies across all samples, rendering them insensitive to these nuanced deception patterns. To address the above issues, we propose the Large Language Model Guided Graph Capsule Network with Local-Enhanced Alignment (LGCNLA) model, a framework that systematically targets three critical dimensions of deception. First, at the semantic layer, we leverage Large Language Models to generate multi-perspective image-enhanced descriptions and supplement external entities, making implicit misleading cues explicit and providing richer contextual evidence. Second, at the structural layer, we construct a dual-channel heterogeneous graph capsule network whose dynamic routing mechanism preserves and amplifies local anomalous patterns that traditional graph based methods would smooth away. Third, at the distributional layer, we reformulate fake news detection as a hierarchical domain alignment task and introduce Local-Enhanced Maximum Mean Discrepancy, which separately aligns cross-modal distributions for real and fake news by minimizing divergence for authentic content while maximizing it for deceptive content, with locality-aware sensitivity to fine-grained inconsistencies. These three components function synergistically such that semantic enrichment provides the evidence, structural preservation highlights the anomalies, and distributional calibration defines the decision objective, thereby creating a holistic detection system that penetrates the camouflage of advanced deceptive strategies. Extensive experiments on Weibo, PolitiFact, and GossipCop datasets demonstrate that our method outperforms the state-of-the-art methods, with accuracy improvements of at least 0.021, 0.029, and 0.036 respectively. The main code is available at https://github.com/sgysgywaityou/LGCNLA.