Malicious Social Bot Detection in Social Networks Based on Multi-modal Feature Fusion with Transformer Networks

With the development and increasing importance of social networks, the threat of malicious social bots in social networks has become increasingly serious. They spread false information, distort public opinion, and damage network security and social stability. Currently, the detection methods for malicious social bots mainly adopt single-modal features, such as text content or user behavior. However, these methods are unable to identify complex and concealed malicious activities. To address these challenges, in the same study, we propose a multi-modal feature fusion method based on Transformer, which combines user behavior, text content information, social network structure, and sentiment features. This method utilizes the Transformer network for multi-modal feature modeling, cross-modal learning, and global context encoding, enabling efficient and robust detection of malicious bots. Experiments show that this method has better detection accuracy and robustness on various social network platforms.