Advances And Challenges In Multimodal Deep Learning For Social Media Fraud And Misinformation Detection

Abstract

The rapid growth of social media has paved the way for the proliferation of deceptive and misleading content, thereby making it essential to develop trustworthy and automated ways. This systematic review is concerned with the evolution of the use of multimodal deep learning methods. For detecting fraud and misinformation on social media, highlighting aspects like datasets modalities ways, performance level, research gaps, and emerging trends. Articles which contain several types of content like text images audio, videos, social media features, and behavioural patterns. Also present a mix of machine learning techniques, deep learning, to advanced models like multimodal fusion attention mechanisms. The evaluation done here has revealed that usually there has been an evolution from traditional single-modal or text-based machine learning to multimodal, contextual, and deep architectures. The social media datasets most extensively utilized in this domain are Twitter, Weibo, and fakeddit. Most multimodal combinations still involve either text-image or text-social-context pairings. Another outcome is the realization that multimodal integration is still quite fragmentary, as is dataset, evaluation, and even multimodal fusion coverage. The survey has pinpointed a research gap that the lack of unified systems that, at the same time, model the content user behaviour social links, external knowledge, and the time context. Because of this, the paper offers a setup for a systematic knowledge of the state of the art and proposes ways forward aimed at developing more holistic, scalable, and trustworthy multimodal detection systems.