A PRISMA based systematic review of deep learning for multimodal fake news detection using text and images

The rapid growth of digitalization has transformed how news is produced and consumed, but it has also accelerated the spread of multimodal fake news, in which misleading text is deliberately combined with manipulated or out-of-context images to deceive audiences. Because unimodal, text-only methods cannot capture cross-modal inconsistencies, deep-learning-based multimodal detection has become a critical research frontier. This systematic literature review, conducted in accordance with the PRISMA 2020 guidelines, synthesizes methods for multimodal fake news detection that combine textual and visual data using deep learning. A structured Boolean search of Scopus and Web of Science retrieved 1,123 records; after duplicate removal, title/abstract screening, and full-text eligibility assessment, 81 studies published between 2020 and 2026 (a partial year) were included. The review spans five analytical dimensions: deep-learning architecture, modality and feature representations, fusion strategies, benchmark datasets, and evaluation metrics. The findings reveal a clear paradigm shift from recurrent and convolutional networks toward pre-trained transformer encoders (BERT, RoBERTa, and Vision Transformers) and graph-based models, and from naive feature concatenation toward attention-driven joint and hybrid fusion. Text-plus-image is the dominant modality combination, while the Weibo and MediaEval Twitter corpora remain the most-used benchmarks, and accuracy and F1-score are the most frequently reported metrics. Rather than presenting this shift as the current frontier, the review positions these findings as a consolidated, reproducible baseline of the text-image deep-learning literature at its transition toward foundation-model-based detection - a record that broad or LVLM-centred surveys presuppose but do not themselves establish. The review further identifies persistent gaps, including modality dominance, over-reliance on internal text–image consistency, limited external verification and explainability, weak generalization under domain shift, dataset and linguistic bias, and high computational cost. Collectively, these findings provide a consolidated foundation and roadmap for future research.