Uncovering Malicious Accounts in Online Social Networks Using XGBoost and Graph Convolution Networks

In this work, we introduce a Machine-Learning (ML) approach based on XGBoost and Graph Convolutional Networks (GCNs) for the detection of active malicious accounts in Online Social Networks (OSN). Distinguishing itself from previous detection algorithms by relying on user relationships, our approach combines heterogeneous information, including descriptive user features and inter-user relationship data, to obtain advanced user representations for determining their legitimacy. Specifically, we employ the XGBoost algorithm to extract crucial features, mitigating the impact of redundant ones. Next, we utilize the Graph Convolutional Networks approach to derive advanced user representations based on the actual relationships between malicious accounts and the descriptive features of users, and lastly, we employ the sigmoid function to assess user legitimacy based on their high-level feature matrices. Experimental results depict that the proposed XGBoost-based approach outperforms other feature selection methods, and comparison with other account detection algorithms shows superior efficiency, accuracy, and flexibility.