A systematic review of multimodal fake news detection on social media using deep learning models

The volume of data circulating from online sources is growing rapidly and comprises both reliable and unreliable information published through many different sources. Researchers are making plausible efforts to develop reliable methods for detecting and eliminating fake web news. Deep learning (DL) methods play a vital role in addressing various fake news detection problems and are found to perform better compared to conventional approaches, making them state-of-the-art in this field. This paper provides a comprehensive review and analysis of existent DL-based models for multimodal fake news detection, focusing on diverse aspects, including user profiles, news content, images, videos, and audio data. This study considered the latest articles within the last seven years, starting from 2018 to 2025, and about 963 quality articles were obtained from the journals and conferences selected for this study. Subsequently, 121 studies were chosen for our SLR after careful screening of the abstract and the full-text eligibility analysis. The findings showed that the Transformer models and Recurrent Neural Networks (RNNs) are the most popular deep learning techniques for detecting multimodal fake news, followed by the Convolutional Neural Networks (CNNs) techniques. The Twitter and Weibo datasets are the two most frequently used standard datasets, and the most frequently used metrics to evaluate the performance of these models are the accuracy, precision, recall, and F-scores. In conclusion, the limitations of the current methods were summarized and some exciting possibilities for future research were highlighted, including designing robust multilingual fake news detection systems, hybridization of deep learning models to enhance detection accuracy, integration of explainable AI (XAI), and facilitating real-time fake news detection models.