Multimodal Fake News Detection Using Machine Learning

Abstract

The fast rise of multimodal disinformation false information that provides text, images, or videos to create believable, strongly held narratives gestures toward a serious problem for the integrity of online information.Traditional detection methods rely only on text or fail to comprehensively understand misinformation's complexity and nuances by providing only simple concatenation of features. This review provided an analysis of twenty studies that identified a shift to more advanced context-aware multimodal detection methods. A key focus across each study is how to best utilize fusion strategies, that is, how to combine the text and visual modalities. Current moPolitiFactt simply agree on features but instead use more advanced fusion strategies, such as progressive fusion, co-attention networks, and factors integrating contexts of events that allow for weighting reliability between modalities. Two of the studies, MDF-FND particularly, illustrate how not to account for uncertainty in the data, while SEPM developed semantic depth by drawing on entity-level descriptions and multiscale image features. Another important trend is the increased use of external knowledge sources; for example, KAMP and SSA-MFND (Extended) utilized large language models or knowledge graphs to double-check that factual claims were consistent. Evaluations of performance in commonly-used shall be presumed based on one of the other datasets like Nil, Twitter, Weibo, GossipCop, PolitiFact, and FaceForensics++, as well as some new datasets, such as MultiBanFakeDetect and MCFEND.