The proliferation of fake news across social media platforms represents one of the most significant challenges to information integrity in the modern digital age. This paper presents a novel multimodal approach to fake news detection that leverages the power of Graph Neural Networks (GNNs) to integrate textual and visual features in a unified framework. Our method constructs rich graph representations of news articles, where nodes represent both textual and visual elements, and edges capture their complex relationships through pointwise mutual information (PMI). We employ an enhanced GraphSAGE architecture with attention mechanisms for effective feature aggregation and classification. Extensive experiments on the Pheme dataset demonstrate that our approach achieves 88.1% accuracy, significantly outperforming existing methods. Through comprehensive ablation studies, we validate the effectiveness of each component in our framework, particularly highlighting the importance of cross-modal feature integration and attention mechanisms. Our method shows particular robustness in handling cases where either modality may be manipulated, offering a promising solution for real-world applications in misinformation detection.