A Multimodal Fake News Detection Model Based on Cross-Image Semantic Fusion

In recent years, social media has become one of the most popular ways of news dissemination. There is a phenomenon that numerous fake news are spreading rampantly on public social media platforms, posing a serious threat to the credibility of social media. Moreover, more and more social media news posts carry multimodal contexts, i.e., utilize not only text but also abundant images to describe the news. However, existing methods only involve the first image along with text in multimodal fake news detection. It severely hampers the extraction of global image semantic information and consequently damages the effectiveness of multimodal fake news detection. To address this issue, we propose a Cross-Image Semantic Fusion based multi-modal fake news detection method (CISF for short). The method uses an adaptive attention diffusion module to model semantic correlations among different images, fully leveraging the contextual dependencies between different images to achieve semantic interaction and fusion among images. On the basis, a global image semantic representation is generated to represent the entire image modality. Finally, the fake news detection is performed based on the multimodal fusion of the text and the global image semantic representation. We conduct experiments on two real-world datasets and demonstrate the effectiveness of the proposed method.