MULTIMODAL FAKE NEWS DETECTION ON SOCIAL MEDIA USING MACHINE LEARNING AND DEEP LEARNING

Abstract

The unchecked propagation of fabricated information across digital networks has emerged as one of the foremost threats to democratic discourse and public confidence in institutions. Existing content-moderation systems relying on manual verification and rigid rule-based filters have proven structurally inadequate against increasingly sophisticated adversarial tactics. This paper presents a multimodal artificial-intelligence framework that simultaneously processes three distinct data streams—user profile metadata, natural-language post content, and associated visual media—to classify social-media content as genuine or fraudulent. A hybrid classification architecture combines Random Forest and feed-forward Neural Network models, augmented by TF-IDF text vectorization, CLIP-based image embeddings, and Optical Character Recognition to surface concealed textual cues embedded in images. The unified feature vector produced by cross-modal fusion yields measurably superior discrimination compared to single-modality baselines. Experimental evaluation against a held-out test partition demonstrates a Random Forest accuracy of 93.8% and a Neural Network accuracy of 95.5%, outperforming prior unimodal approaches by a statistically significant margin. A lightweight Flask-based inference endpoint enables sub-two-second real-time predictions suitable for platform integration. The results confirm that cross-modal feature fusion is an effective and scalable strategy for combating the modern fake-content ecosystem.