Practical Privacy-Preserving Community Detection in Decentralized Weighted Networks

Tingxuan Han, Wei Tong, Jiacheng Niu, Sheng Zhong

Abstract: As one of the essential graph analysis tasks, community detection plays a crucial role in various applications. Since a graph can be viewed as a collection of users’ personal information and the relationships among them, it is essential to protect individuals’ privacy during the process of community detection. Weighted networks are networks with real-valued weights on edges, and the edge weights should be protected when performing community detection in weighted networks. In this paper, we study the privacy-preserving community detection problem for weighted networks where each node is a distributed user and there is no trusted third party. Directly applying homomorphic encryption to implement a secure version of community detection for the untrusted analyst will incur high communication and computational cost, while local perturbation methods, such as local differential privacy, may significantly degrade the accuracy. Therefore, we combine these two types of techniques and propose a practical privacy-preserving algorithm for community detection in weighted networks, by blending a tailored locally differentially private mechanism with a cryptographic component. The proposed method can provide provable privacy guarantees and satisfactory performance. Extensive experiments we conducted have demonstrated the accuracy and efficiency of the proposed algorithm in various cases.