The automated detection of fake news has become challenging in the light of the spread of misinformation on social media, particularly when it involves deciphering the multiple modes of data, including textual information, audio information and visual information. By only relying on traditional uni modal methods, inconsistencies and complementary relationships between these modalities are not usually recognized. In this study, a Multimodal Global and Regional Transformer Network (MGRTN) is proposed to detect fake news and deal with their binary classification. The framework relies on the use of a pretrained language model that can extract semantic information from the linguistic aspect of news and a visual feature extraction module to capture both the overall characteristics of the image and the local ones. The multimodal representations are fused based on interaction with transformer-based learning, which enables the model to learn the relationship between textual and visual evidence before the final classification. Evaluation on the Fakeddit multimodal dataset shows that the proposed approach gives an accuracy of 91.88%, a precision of 93.48%, a recall of 93.03% and an F1-score of 93.25%. The results show that using multi-level visual evidence and contextual language representations together led to effective multimodal representations for identifying misleading news content. The proposed framework provides a basis for creating more powerful automated misinformation detection systems in multimodal social media.

multimodal learning,deep learning,Natural Language Processing,Terms—fake news detection,multimodal transformers,computer vision,fakeddit,multimodal classification