差分隐私在数据保护中的应用

摘要：针对匿名化、脱敏等传统技术难以抵御重新识别与差分攻击的局限，提出一种融合中心化与本地化差分隐私的通用数据保护架构，可灵活配置两种模式：本地化模式下由终端完成数据扰动，保障原始数据不出域；中心化模式下由数据管理器统一加噪，实现隐私泄露的量化管控。同时明确3类组件的功能要求，并结合具体实践验证其可行性。差分隐私可有效解决传统方案对攻击者背景知识的依赖，为数据“可用不可见”提供可落地的技术路径。

关键词：差分隐私；数据安全；隐私保护

Abstract: To address the limitations of traditional techniques such as anonymization and desensitization in resisting re-identification and differential attacks, a general data protection architecture that integrates centralized and local differential privacy is proposed, supporting flexible configuration of two modes: in the local mode, data perturbation is performed at the terminal to ensure that raw data does not leave its domain; in the centralized mode, noise is added uniformly by the data manager to achieve quantitative control over privacy leakage. The functional requirements of the three types of components are also specified, and the feasibility of the proposed architecture is validated through practical implementations. Differential privacy can effectively eliminate the dependence of traditional schemes on attackers' background knowledge and provide a practical technical pathway for making data "usable but invisible."

Keywords: differential privacy; data security; privacy protection