DSFENet: A Domain-Specific Feature-Enhanced Generalizable Model for Multimodal Fake News Detection

The diverse and emerging fake news on social media severely threatens the safety of society, so it is crucial to timely curb fake news. However, the existence of domain discrepancy makes traditional deep learning-based detection methods perform poorly on unseen domains. Despite existing works reducing domain discrepancy, most of them only focus on domain-invariant features while ignoring the importance of domain-specific features, and rely on target domain data during training. Therefore, we propose a domain-specific feature-enhanced detection network (DSFENet) for multimodal fake news detection, which achieves domain generalization by jointly modeling domain-invariant features and domain-specific features in a unified model. Specifically, we leverage multimodal supervised contrastive learning on domain-invariant features to reduce domain discrepancy. Then, we impose mutual information estimation between domain-specific and domain-invariant features to increase the feature diversity for better generalization. Finally, based on the complementary guidance nature of multimodal information, we design a cross-modal translation module to regularize the learned features. Experimental results on public multimodal datasets Pheme and Twitter demonstrate the superiority of the DSFENet.