Improved Privacy-Preserving Clustering Model for Distributed Data
TL;DR
A CKKS-based distributed k-means method that represents each centroid as a packed ciphertext, indicating that centroid packing reduced operational cost without adding clustering-quality loss.
Abstract
Privacy-preserving distributed clustering lets organisations analyse data held in separate locations without centralising raw records. Homomorphic encryption can protect centroid exchange, but distributed k-means becomes slow and communication-intensive when every centroid feature is encrypted and transmitted separately. This article presents a CKKS-based distributed k-means method that represents each centroid as a packed ciphertext. Under a semi-honest coordinator assumption, participant records remain local, centroid updates are encrypted before transmission, and aggregation is performed on ciphertexts. The method was evaluated on Iris, Wine, Diabetes, and Credit Card Default datasets under simulated 3-participant and 5-participant settings. Three configurations were compared: centralised plaintext k-means, a per-dimension CKKS encrypted centroid baseline, and the packed CKKS method. Clustering quality was measured using silhouette score and Davies-Bouldin index, while efficiency was measured using encryption time, total execution time, and encrypted communication volume. Compared with the per-dimension encrypted baseline, the packed method reduced encryption time by 57.47% to 95.27% and encrypted communication volume by 75.00% to 95.65%. In the Credit Card Default 3-participant setting, total runtime fell from 24.08 seconds to 1.12 seconds, and encrypted communication fell from 450,504.18 KB to 19,587.03 KB. The silhouette-score difference relative to the encrypted baseline was 0.0000 in all reported encrypted comparisons, indicating that centroid packing reduced operational cost without adding clustering-quality loss.