Scholarlink

<p>Many applications, such as traffic monitoring and usage analytics, need to collect user data repeatedly over time. Local Differential Privacy (LDP) protects individual records before they leave the user’s device, but existing LDP mechanisms for continual reporting face two practical limitations: (i) they offer a single, fixed privacy level for all users, forcing the server to adopt the most conservative budget and reducing accuracy; and (ii) grouping users by privacy preference, as in personalized schemes, shrinks the effective population at each level, which again hurts utility. This paper introduces <b>S</b>treaming and <b>Hi</b>ghly-personaliz<b>e</b>d <b>L</b>ocal <b>D</b>ifferential Privacy (SHieLD), a modular framework that addresses both issues. On the client side, SHieLD adopts difference-tree structures to efficiently report changes in sequential data under each user’s chosen privacy budget. On the server side, it contributes two mechanism-independent components: a <i>sampling-based Data Replication</i> process that amplifies the effective population at stricter privacy levels by sub-sampling data from less restrictive ones, and a <i>Combination</i> process that merges frequency estimations across levels through optimized weighted aggregation. Unlike earlier noise-based replication techniques, our sampling approach adds no extra noise and can be used with any LDP algorithm. We evaluate SHieLD on four datasets (two Gaussian, one uniform, one constant) across multiple privacy configurations with population sizes of 10&#xa0;000 and 50&#xa0;000 users. Compared with RAPPOR and DDRM, SHieLD consistently achieves lower Mean Squared Error and Mean Absolute Error, particularly when personalization is enabled. The experiments also confirm that Data Replication and Combination each contribute independently to accuracy gains.</p>

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

SHieLD: a personalized local differential privacy framework for continual reports

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月14日

查看全文
  • Majid Zolfaghari,
  • Ahmad Mohammadi,
  • Rasool Jalili

摘要

Many applications, such as traffic monitoring and usage analytics, need to collect user data repeatedly over time. Local Differential Privacy (LDP) protects individual records before they leave the user’s device, but existing LDP mechanisms for continual reporting face two practical limitations: (i) they offer a single, fixed privacy level for all users, forcing the server to adopt the most conservative budget and reducing accuracy; and (ii) grouping users by privacy preference, as in personalized schemes, shrinks the effective population at each level, which again hurts utility. This paper introduces Streaming and Highly-personalized Local Differential Privacy (SHieLD), a modular framework that addresses both issues. On the client side, SHieLD adopts difference-tree structures to efficiently report changes in sequential data under each user’s chosen privacy budget. On the server side, it contributes two mechanism-independent components: a sampling-based Data Replication process that amplifies the effective population at stricter privacy levels by sub-sampling data from less restrictive ones, and a Combination process that merges frequency estimations across levels through optimized weighted aggregation. Unlike earlier noise-based replication techniques, our sampling approach adds no extra noise and can be used with any LDP algorithm. We evaluate SHieLD on four datasets (two Gaussian, one uniform, one constant) across multiple privacy configurations with population sizes of 10 000 and 50 000 users. Compared with RAPPOR and DDRM, SHieLD consistently achieves lower Mean Squared Error and Mean Absolute Error, particularly when personalization is enabled. The experiments also confirm that Data Replication and Combination each contribute independently to accuracy gains.