Differentially-private Multiparty Clustering

Abstract

In an era marked by the widespread application of Machine Learning (ML) across diverse domains, the necessity of privacy-preserving techniques has become paramount. The Euclidean k-Means problem, a fundamental component of unsupervised learning, brings to light this privacy challenge, especially in federated contexts. Existing Federated approaches utilizing Secure Multiparty Computation (SMPC) or Homomorphic Encryption (HE) techniques, although promising, suffer from substantial overheads and do not offer output privacy. At the same time, differentially private k-Means algorithms fall short in federated settings. Recognizing the critical need for innovative solutions safeguarding privacy, this work pioneers integrating Differential Privacy (DP) into federated k-Means. The key contributions of this dissertation include the novel integration of DP in horizontally-federated k-Means, a lightweight aggregation protocol offering three orders of magnitude speedup over other multiparty approaches, the application of cluster-size constraints in DP k-Means to enhance state-of-the-art utility, and a meticulous examination of various aggregation methods in the protocol. Unlike traditional privacy-preserving approaches, our innovative design results in a faster, more private, and more accurate solution, significantly advancing the state-of-the-art in privacy-preserving machine learning.