TriSageNet: A Multimodal Attention-Guided Framework for High-Precision Fake News Detection

The multiform nature of internet misinformation that undermines public discourse calls for powerful and multifunction multimodal detection systems. Existing techniques are often based on unimodal analysis or simple fusion techniques, which limit the capturing of complex textual, visual, and relational information. This paper presents TriSageNet, a new tri-stream, attention-based architecture for the accurate detection of synthetic news. TriSageNet applies semantic encoding of textual information, visual information from images into structured visual representation, and graph-based relational representation among information sources for measuring source credibility. A hierarchical attention architecture is adopted, which adaptively incorporates these modalities, focusing on contextually salient features calculated in various categories of news. Experiments were conducted with a benchmark dataset called Fakeddit, containing rigorously annotated multimodal news articles. TriSageNet significantly outperformed five major baseline models—namely CNNs, BERT+BILSTM, BERT+BIGRU, RoBERTa+GNN+HAN, and GPT-4 Omni—with an accuracy of 95.89%, precision of 94.88%, recall of 93.9%, and an F1 score of 93.89%. These results highlight robust improvement in reducing error rates, modality synergy, and generalization ability. The proposed approach provides a scalable and technically significant benchmark for misinformation detection and demonstrates high potential for deployment in high-performance dynamic digital applications.