Differentially Private Graph Clustering Algorithm Based on Structure Similarity

Abstract

With the widespread use of new information systems such as social networks, recommendation systems as well as location-based services, graph data has become a very common and important data type. It has been shown that, from these collected graph data, some special substructures can be found through clustering analysis, and can further support the intelligent decision. However, directly publishing or using clustering results on these graph data would disclose the privacy information of system users. To this end, based a classical structural clustering algorithm for networks (SCAN) and the technology of differential privacy, we propose a differentially private graph clustering algorithm named DP-SCAN. Specifically, we first reasonably calibrate the global sensitivity of the function of computing structure similarity between nodes in the graph, and thus specify parameters of the Laplace mechanism for capturing differential privacy. Then, we provide details of the DP-SCAN algorithm. The theoretical analysis indicates that DPSCAN algorithm satisfies ε-differential privacy, without trading off the clustering efficiency. The experimental results show that, when compared with the original SCAN clustering algorithm, DP-SCAN clustering algorithm can maintain the validity of clustering under the premise of satisfying differential privacy.