A Deep Learning Approach to Fake News Detection in Images with Text Overlay

Abstract

This study addresses the growing challenge of fake news distributed through images containing overlaid text, a format that is widely used on social media because it is visually persuasive, easy to share, and often difficult to verify. Unlike conventional text-only misinformation, this type of content combines visual and textual cues that can strengthen misleading narratives and increase public trust in false information. To respond to this problem, the study proposes a deep learning approach for detecting fake news in text-overlay images by integrating image-based and text-based feature extraction within a unified classification framework. The proposed method begins with image preprocessing and text extraction from embedded captions, followed by feature learning using deep neural architectures to capture both semantic and visual patterns associated with deceptive content. The model is trained and evaluated on a labeled dataset of images containing news-like textual overlays, with performance assessed using accuracy, precision, recall, F1-score, and confusion matrix analysis. The results indicate that the proposed approach is able to identify fake news content effectively and achieves promising classification performance compared with baseline machine learning methods. These findings indicate that combining visual representation and embedded textual information can significantly improve detection capability in multimodal misinformation settings. This study aids in the advancement of more flexible fake news detection systems, especially in digital contexts where altered or deceptive image-based content disseminates swiftly. The proposed framework is expected to support future research and practical implementation in automated content verification, social media monitoring, and digital information integrity management