Beyond Single-Modality Detection: A Systematic Review of Multimodal AI for Social Media Cybercrime

Abstract

Social media platforms have emerged as a key battleground for cy-bercriminals, which leverages the speed, scale and interconnectedness of social media to orchestrate financial crimes such as phishing and OTP Fraud, identity-based crime such as deep-fake and account takeover, and social harms such as cyberbullying and coordinated influence campaigns. The threats are increasingly multifaceted and cross-platform, making it difficult for traditional detection solutions to stay ahead of the curve, as they are typically limited to a specific data type or a single platform. We review recent publications on detecting cybercrime in social networks ranging from misinformation detection, to machine learning and deep learning based classifiers, fraud and intrusion detection pipelines, to novel multimodal architectures for AI. The reviewed literature can be divided into three the-matic groups: (1) detection methods based on content and propagation patterns, (2) methods based on user behaviour and network structure, and (3) multimodal fusion models coupled with explainable AI mechanisms. We evaluate the data sets used, algorithm selection, reported performance, and inherent limitations of the underlying studies for each cluster. The analysis shows that there is a gap. Although there has been significant progress in the detection of individual tasks, existing systems are still highly specialised and platform-dependent, or type-dependent, or modal-ity dependent. Very few integrate text, image and behaviour into a single model. It also identifies a lack of attention to interpretability and cross-platform real time operation. To address these gaps, we suggest the necessity of developing a multimodal detection system that is uniform, adaptive and explainable, which can work in heterogeneous environments on social media platforms, and we propose some research directions for achieving this goal.