Fake News Detection with Multimodal Interaction and Unsupervised Learning Fusion

The swift advancement of generative artificial intelligence greatly enhances the spread of disinformation, facilitating quick learning from a minimal set of annotated examples. This study presents a fake news detection model (MIUSL) that combines different types of data and unsupervised learning, which helps to improve accuracy in detecting false news and lowers the costs of manual labeling by using text, images, and sentiment information. The process includes pulling out text and emotion details using BERT, getting image details with CNN, combining these different types of features using a multi-head attention method, and using a contrastive loss function to better align the features; it then merges three unsupervised models to make the final detection through a weighted voting system. The studies use the microbomb multimodal dataset, and the results show that the MIUSL model is more accurate than the 11 baseline models. The tests show that taking away any part of the multimodal data causes a big drop in performance, proving that using multiple types of data together is effective. This paper's innovation is in combining sentiment features for better matching of text and image feelings, and using an unsupervised model to enhance detection variety, which greatly lowers the need for manual labeling and offers a smart, budget-friendly way to spot false information.