<p>With the rapid development of social media, the generation and dissemination of fake news have become increasingly convenient, making fake news detection a widespread challenge in the field of information dissemination. Existing non-graph-based deep learning methods often struggle to effectively model complex relationships across multimodal data. In contrast, Graph Neural Networks (GNNs) can model multimodal data as graph structures to better capture intricate intermodal relationships, though incurring substantial computational overhead. To address these challenges, we propose the Interactive Graph Network Compression (IGNC) model. This model employs a fine-grained graph construction module to extract rich unimodal feature representations and construct unimodal graph structures. It then utilizes a semantic-based graph compression module to eliminate semantically redundant connections. Finally, through a multimodal graph interaction module, it achieves a stable transformation from unimodal to multimodal graph structures. By enhancing intra-modal and inter-modal information interaction through attention mechanisms, the model learns more robust news representations. Experiments on public datasets demonstrate that IGNC significantly improves fake news detection accuracy while effectively reducing semantic redundancy.</p>