MACSD: Multimodal Spammer Detection Based on Multi-Path Cascaded Auto-Encoder

Online social networking platforms provide communication channels for users worldwide. Meanwhile, the behavior of spammer groups quickly spread worldwide and caused serious harm to cyberspace. Currently, spammer identification research mainly focuses on the unimodal modeling of user behavior. However, spammer behavior usually includes multimodal evidence to support their misleading statements. Therefore, we propose a spammer-detection model that integrates multimodal behavioral features and key time node analysis. First, a pre-trained model is used to mine unimodal user behavior features. Second, the association relationship between multimodal behavioral features is deeply mined by integrating the multi-head attention component (MHA) and the auto-encoder component. Finally, considering the indirect temporal coherence and suddenness of user group behavior, historical behavioral sequences are encoded using absolute position encoding (APE). Moreover, the attention mechanism is combined to analyze temporal behavioral features to capture the key time nodes and identify the spammer accounts. Experiments on the public WEIBO and TWITTER datasets show that MACSD achieves competitive and, in many settings, superior performance compared with strong baseline methods.