Improving sybil detection in online social networks

Abstract

Due to open and anonymous nature, online social networks are particularly vulnerable to Sybil attack, where a malicious user can fabricate many dummy identities to target systems. Recently, there are a flurry of interests to leverage social network structure for Sybil defense. The rationale behind these approaches is to partition the whole social graph into honest and Sybil regions based on two core assumptions: (1) strong trust relationship among nodes, which makes it difficult for Sybil nodes to establish substantial social connections with non-Sybil nodes, even if they can easily recruit a large number of Sybil nodes and build an arbitrary topology network among them. As a result, Sybil region connects to the main network via a small number of attack edges. (2) honest region is fast mixing, where random walks from an benign node can quickly reach a stationary distribution after O(log(n)) steps, compared to random walks from Sybil nodes. However, these anti-Sybil mechanisms are less effective than expected since real-world social networks do not conform to the above assumptions. In the thesis, the main theme is to explore additional topological features underlying social networks in designing secure and robust Sybil defense systems. In particular, this thesis is comprised of two studies. Firstly, graph pruning and regularization techniques are proposed to enhance existing network-based Sybil defense mechanisms. In particular, a novel perspective is provided to interpret Sybil defense as the problem of partially labelled classification. Then, based on this framework, graph pruning is introduced to enhance the robustness of current anti-Sybil schemes against target attacks, by utilizing the local structural similarity between neighboring nodes in a social network. Besides, a domain-specific graph regularization method is designed to further improve the performance of those mechanisms by exploiting the relational property of social networks. Experimental results on four popular online social network datasets demonstrate that the proposed techniques can significantly improve the detection accuracy over the original Sybil defense mechanisms. i I would like to express my sincere thanks to my supervisor Prof. Jie Zhang, and Prof. Chang Kuiyu from School of Computer Engineering, who have continuously encouraged, inspired and supported me throughout my master research. I would also like to thank all my friends and staffs in Center for Computational Intelligence Lab (C2I) and Data-Intensive Scalable Computing Lab (DISCO) for their technical support and enthusiastic help. Many thanks are given to the members of the research group I am working in, particularly Chang Xu, Hui Fang, Yuan Liu, Guibing Guo, Siwei Jiang. Without them, the research cannot be such an enriching and enjoyable experience. My immense gratitude goes to my family for their great love and continuous encouragement in both my studies and life.