Scholarlink

As artificial intelligence advances, data clustering faces significant privacy challenges. Existing Differential Privacy (DP) k-modes methods typically rely on fully trusted data collectors, while Local Differential Privacy (LDP) underperforms compared to DP; neither approach adequately addresses personalized privacy requirements. This paper introduces a novel Personalized Local Differential Privacy k-modes (PLDP k-modes) algorithm. This method utilizes a PK-RR mechanism to perturb data locally at the user side and employs server-coordinated iterative centroid updates, thereby protecting users’ real data and respecting personalized privacy needs. A key contribution includes the incorporation of a weight function into the perturbation algorithm to enhance utility, alongside a centroid perturbation method to counteract inference attacks. Notably, this marks the first application of PLDP in k-modes clustering. Theoretical analysis and experimental results confirm the algorithm’s privacy guarantees and demonstrate a superior privacy-utility tradeoff, with its clustering utility surpassing that of state-of-the-art DP k-modes algorithms. The proposed method effectively addresses the privacy-utility challenge in k-modes clustering.

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Privacy-Preserving Framework for k-Modes Clustering Based on Personalized Local Differential Privacy

  • 会议论文

  • 发表日期：

    2026年

查看全文
  • Yuling Luo,
  • Zhangrui Wang,
  • Xue Ouyang,
  • Siyuan Zu,
  • Qiang Fu,
  • Sheng Qin,
  • Junxiu Liu

摘要

As artificial intelligence advances, data clustering faces significant privacy challenges. Existing Differential Privacy (DP) k-modes methods typically rely on fully trusted data collectors, while Local Differential Privacy (LDP) underperforms compared to DP; neither approach adequately addresses personalized privacy requirements. This paper introduces a novel Personalized Local Differential Privacy k-modes (PLDP k-modes) algorithm. This method utilizes a PK-RR mechanism to perturb data locally at the user side and employs server-coordinated iterative centroid updates, thereby protecting users’ real data and respecting personalized privacy needs. A key contribution includes the incorporation of a weight function into the perturbation algorithm to enhance utility, alongside a centroid perturbation method to counteract inference attacks. Notably, this marks the first application of PLDP in k-modes clustering. Theoretical analysis and experimental results confirm the algorithm’s privacy guarantees and demonstrate a superior privacy-utility tradeoff, with its clustering utility surpassing that of state-of-the-art DP k-modes algorithms. The proposed method effectively addresses the privacy-utility challenge in k-modes clustering.