Title: PE-means: Enhancing Differentially Private $k$-means Clustering via Private Evolution

Original: arXiv:2606.00342v1 Announce Type: new

Abstract: This research investigates the challenge of performing differentially private (DP) $k$-means clustering within Euclidean space. Existing approaches typically involve aggregating private data directly, a process that results in sensitivity levels scaling with the size of the domain. To address this, we present PE-means, which adapts the private evolution (PE) algorithm—a method gaining traction for synthetic data generation—to the specific context of $k$-means clustering. A primary benefit of the PE framework is its reliance on computing a private histogram that maintains constant sensitivity, thereby directing the evolutionary process. Our implementation introduces novel evolutionary operators tailored for clustering tasks, alongside other algorithmic enhancements that hold broader significance. Empirical results demonstrate that PE-means delivers an average reduction of 20% in clustering loss compared to current state-of-the-art baseline methods.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC