Practical and Robust Privacy-Preserving K-Means Clustering Protocols With GPU Acceleration

Privacy-preserving machine learning allows for the protection of the original data's privacy while training models. In recent years, researchers have implemented machine learning algorithms using secure multi-party computation (MPC) techniques, with clustering being one of the key applications. However, current privacy-preserving clustering algorithms still face challenges regarding execution speed and communication overhead. In this paper, we propose a new three-party GPU-accelerated privacy-preserving K-means clustering algorithm aimed at improving computational efficiency in distributed settings. We design a three-party K-means algorithm and optimize subprotocols such as the truncation protocol and the constant-round comparison protocol to reduce communication latency. Additionally, we leverage the powerful capabilities of Graphics Processing Units (GPUs) to enhance the execution efficiency of the algorithm, particularly for intensive distance computation and centroid updates. Finally, we propose a privacy-preserving clustering quality evaluation protocol to address the parameter selection issue (determining the optimal number of clustersK) in privacy-preserving K-means. Experimental results on the synthetic datasets demonstrate that our method achieves a137×–221×improvement in efficiency compared to the CPU-based protocol of Mohasselet al., and a5.7×–11.2×speedup over the CPU-based state-of-the-art implementation by Weiet al..