<p>Since the adverse impact of fake news, especially multi-modal fake news, on public decision-making and social governance, multi-modal fake news detection has lately attracted increasing attention. However, many existing methods ultimately exploit multi-modal features without detail fusion to complete detection and insufficiently consider the intrinsic features in news content, resulting in poor performance. To tackle these issues, we propose a network for multi-modal fake news detection that uses hierarchical multi-grained features fused with global latent topic (HMLTNet). Specifically, we first construct Hierarchical Multi-grained Encoding Module to capture convolutional and hierarchical textual features. Then, Cross-modal Shared Attention Module completes detail compensation in the multi-modal features by fusing textual and visual features and jointly modeling inter- and intra-modality correlations. Finally, the global latent topic features are excavated and stocked from multi-modal features by utilizing Latent Topic Memory Module. Furthermore, we design an Enhanced Similarity Module and introduce a dense-like strategy together to alleviate the adverse effects of cross-modal semantic gap. Extensive experiments on three public datasets indicate that the presented network reaches the best accuracy compared to state-of-the-art methods.</p>