DEMFND: Domain-Enhanced Multimodal Fake News Detection

The proliferation of multimodal fake news presents a significant challenge, demanding robust detection techniques. However, existing methods face the following limitations: (1) inadequate capture of deep cross-modal semantic relationships within specific domains, (2) poor generalization across diverse news domains, and (3) difficulty integrating multi-granularity features while mitigating domain biases amplified by pre-trained models. To address these issues, in this paper we propose DEMFND, a novel Domain-Enhanced Multimodal Fake News Detection framework. Specifically, DEMFND first extracts multi-granularity features using BERT, MAE, and CLIP. Subsequently, we employ a domain-aware multi-expert feature enhancement layer to learn domain-specific characteristics, incorporate domain-aware multimodal contrastive learning for domain-conditioned semantic alignment, and utilize a domain-aware multimodal fusion mechanism with gating for adaptive integration in our DEMFND. Comparative experiments conducted on three real-world datasets (encompassing both Chinese and English) against multiple baseline models indicate that our DEMFND outperforms the best baseline (MMDFND) with improvements of 0.64% and 0.96% in accuracy and F1-score, respectively.