Worries about the increasing number of Sybils in online social networks (OSNs) are amplified by a range of security issues; thus, Sybil detection has become an urgent real-world problem. Lightweight and limited data-friendly, LBP (Loopy Belief Propagation)-based Sybil-detection methods on the social graph are extensively adopted. However, existing LBP-based methods that do not utilize node attributes often assume a global or predefined homophily strength of edges in the social graph, while different user’s discrimination and preferences may vary, resulting in local homogeneity differences. Another issue is that the existing message-passing paradigm uses the same edge potential when propagating belief to both sides of a directed edge, which does not agree with the trust interaction in one-way social relationships. To bridge these gaps, we present SybilHP, a Sybil-detection method optimized for directed social networks with adaptive homophily prediction. Specifically, we incorporate an iteratively updated edge homophily estimation into the belief propagation to better adapt to the personal preferences of real-world social network users. Moreover, we endow message passing on edges with directionality by a direction-sensitive potential function design. As a result, SybilHP can better capture the local homophily and direction pattern in real-world social networks. Experiments show that SybilHP works with high detection accuracy on synthesized and real-world social graphs. Compared with various state-of-the-art graph-based methods on a large-scale Twitter dataset, SybilHP substantially outperforms existing methods.

The proliferation of social bots on social networks presents significant challenges to network security due to their malicious activities. While graph neural network models have shown promise in detecting social bots, acquiring a large number of high-quality annotated accounts remains challenging, impacting bot detection performance. To address this issue, we introduce BotCL, a social bot detection model that employs contrastive learning through data augmentation. Initially, we build a directed graph based on follower relationships, utilizing semantic, attribute, and structural features of accounts as initial node features. We then simulate account behaviors within the social network and apply two data augmentation techniques to generate multiple views of the directed graph. Subsequently, we encode the generated views using relational graph convolutional networks, achieving maximum homogeneity in node representations by minimizing the contrastive loss. Finally, node labels are predicted using Softmax. The proposed method augments data based on its distribution, showcasing robustness to noise. Extensive experimental results on the Twibot-20 and Twibot-22 datasets demonstrate that our approach surpasses state-of-the-art methods in terms of performance.