FMC: Multimodal fake news detection based on multi-granularity feature fusion and contrastive learning

The automatic detection of multimodal fake news has recently garnered significant attention. However, the existing detection methods mainly focus on merging textual and visual features, but fail to make full use of multimodal numbers from the perspective of multi-granularity. Additionally, emerging pre-trained multimodal learning models and powerful contrastive learning methodologies remain underutilized in this domain. To address these challenges, we introduce a multimodal fake news detection framework (FMC) that integrates multi-granularity feature fusion with contrastive learning. Initially, FMC utilizes a range of pre-trained models to extract textual and visual features at various levels of granularity. Subsequently, a multi-granularity fused multimodal news representation is created through cross-modal alignment and textual–visual co-attention . The final step involves classifying the news as true or fake, leveraging a combination of multi-head self-attention mechanisms and a contrastive learning auxiliary task. This contrastive learning auxiliary task specifically aims to minimize the distance between similar news representations that share the same label in the multimodal feature space, and maximize the distance between representations of differing labels. Comprehensive experiments conducted on three real-world datasets have demonstrated the superior effectiveness of our proposed framework, significantly surpassing the performance of current state-of-the-art methods.