In social media, the proliferation of multimedia content has amplified the spread of fake news. Various algorithms, from manual feature extraction to deep learning approaches, have been developed to detect fake news, but they often face limitations such as incomplete modal representation, shallow multimodal fusion, and insufficient consideration of modality inconsistency. To address these challenges, this paper introduces the Granularity Enhanced Multimodal High-Order Interaction Network (GMHIN). GMHIN uses a multi-level encoding network to capture detailed and rich semantics and a high-order fusion network with a cross-modal bilinear attention interaction block to effectively fuse images and text. It also employs a feature interaction matrix to ensure consistent representations across modalities. This framework helps identify and mitigate false rumors in multimedia content on social media, thereby reducing misinformation. Extensive experiments demonstrate the effectiveness of GMHIN, showing an improvement in performance by up to 2.1% across two datasets.