Tri-FusionDet: Leveraging User Engagement, Textual, and Visual Features for Enhanced Fake News Detection

The exponential rise of fake news presents numerous social problems, diminishing trust in news sources, intentionally distributing incorrect information to the public, and disrupting social cohesion. Generative Artificial Intelligence (GenAI) has worsened this problem by producing highly realistic fake news through text, images, and other modalities, often indistinguishable from authentic content. To address this pressing and critical issue, advanced multimodal detection methods are crucial. Existing multimodal fake news detection primarily focuses on textual and visual features, neglecting valuable social engagement metrics such as comments, likes, and shares. This hampers the effective differentiation between authentic and manufactured content.In this paper, the Tri-Modal Fusion Detector (Tri-FusionDet), a more sophisticated system for improved multimodal fake news detection, is proposed. It applies a late fusion technique to combine textual analysis, visual processing, and social engagement cues within one framework in order to provide a better assessment of information credibility. This approach marks the first integration of textual, visual, and engagement features for fake news detection using a late fusion method. A comprehensive evaluation of the large-scale Fakeddit dataset demonstrates that Tri-FusionDet surpasses state-of-the-art methods, achieving an accuracy of 94%, which is 6% higher than the best-performing baseline model.