Node Similarity-Based Synthetic Graphs Releasing with Differential Privacy Guarantees

Bin Cai, Yangrui Li, Chunqiang Hu, Weihong Sheng, Jiajun Chen, Jiguo Yu

Abstract: Social graphs are essential for various data analysis, yet their direct publication introduces significant privacy risks due to the sensitive nature of their information. To mitigate these risks, many algorithms employ synthetic graph generation combined with differential privacy techniques for secure data publishing. However, existing approaches often suffer from limitations such as reduced accuracy in representing the original graph, diminished usability, and insufficient privacy protection. To address these issues, we propose a novel node similarity-based DP method, SimDP-CL, specifically tailored for synthetic graph generation of social networks. Our approach enhances node similarity computation by integrating shortest path distances and ensures robust privacy for node similarity and degree distribution through differential privacy mechanisms. By incorporating node similarity from the original graphs and integrating neighbor information from the synthetic graphs, our method probabilistically assigns edges, thereby preserving crucial properties such as degree distribution and global clustering coefficients. Experimental results show that this approach not only maintains the structural integrity of the synthetic graphs but also improves their overall usability, offering a more effective balance between privacy protection and data utility.

External IDs:doi:10.1109/tnse.2025.3597347