A Multimodal Deep Learning Architecture for Scalable Fake News Detection in Online Social Networks

The accessibility of social media platforms has made them a major vehicle for the dissemination and virality of fake news, leading to significant social, political, and economic consequences. Detecting fake news therefore requires automated mechanisms that are both scalable and robust. Most existing systems rely on unimodal textual analysis, assuming misinformation can be identified solely through text. However, in real-world scenarios, misinformation typically involves an interplay of textual, visual, and social components. This paper proposes a multimodal deep learning framework that integrates features from textual, visual, and social context components. The fusion of these modalities is expected to enhance scalability and reliability in fake news detection. Transformer-based models are employed to obtain effective textual representations, while convolutional neural networks (CNNs) are used to capture visual features. Social context is modeled using user engagement and dissemination patterns. Experimental evaluations based on robustness, accuracy, F-score, and other established benchmarks for social media systems demonstrate that the proposed architecture outperforms existing unimodal and multimodal detection models. Furthermore, the framework shows strong scalability in large-scale social media environments. Overall, the study presents a practically deployable and effective solution for detecting fake news and mitigating the spread of misinformation on real-world social media platforms.