Federated Graph Analytics with Differential Privacy

Shang Liu, YANG CAO, Masatoshi Yoshikawa

Keywords: Federated Analytics, Graph Statistics, Differential Privacy

TL;DR: We first present federated graph analytics (FGA), a new paradigm that calculates common graph statistics across several distributed clients.

Abstract: Graph analytics across various sources yields valuable insights; however, ensuring privacy becomes increasingly challenging. Federated analytics offers a promising privacy-preserving framework for data analytics. Nonetheless, most current methods are geared towards generic tabular data, thereby limiting their effectiveness in complex graph analytics. In this paper, we first present federated graph analytics (FGA), a new paradigm that calculates common graph statistics across several distributed subgraphs. A key challenge with FGA is the limited view each client has of the global graph, making it challenging for each client to obtain accurate graph statistics. To tackle this, we propose an FGA framework that supports various graph analytics while providing the privacy guarantee. We find that the overlapping information among distributed subgraphs leads to redundant randomization, influencing the utility significantly. To mitigate this issue, we put forward an improved method based on private set intersection (PSI) techniques. By perturbing each item in disjoint subgraphs only once, we reduce the level of added noise greatly. Extensive experiments conducted on real-world graphs demonstrate that our improved method outperforms the baseline by over 70% in most cases.