面向个性化需求的终端差分隐私保护新方法

Applications built on personal data can create tremendous social and economic value, meanwhile, the extensive collection and analysis bring serious threats to individual privacy. How to make a balance between the exploration of personal data and the protection of individual privacy protection is an urgent issue for academia, industry and government. As a rigorous model of privacy, differential privacy has attracted much attention in recent years. A differentially private algorithm, which takes a dataset as input, can guarantee that the output of the algorithm will not disclose any individual’s information, yet the output is still useful. Current research on differential privacy are usually based on the assumption that there is a trustworthy data collector who can access individuals’ data and conduct data perturbation. Unfortunately, such an assumption hardly holds in practice. Considering this, in this project we plan to investigate the “local” model of differential privacy, where data are first perturbed by individuals and then analyzed by a possibly untrustworthy data collector. On one hand, we will investigate which factors affect individual’s preferences for privacy protection. And by conducting qualitative and quantitative analysis of the relationship among the degree of data perturbation, the degree of privacy protection, and the correctness of data analysis result, we will study how to find the optimal data perturbation strategy for an individual. On the other hand, we will study local differential privacy mechanisms that can be applied to similarity computation. Different individuals usually have different requirements for privacy. As a result, individuals will perturb their data to different extent. We will investigate how to design data analysis methods to deal with this problem. And we will study how to improve the quality of the data analysis result. By conducting above research, we expect to find feasible and effective data perturbation methods and data analysis methods for individuals and data collectors respectively, so that we can provide necessary theoretical, methodological and technical foundations for practical application of differential privacy.

基于个人信息的大数据应用可创造巨大的价值，但个人数据的广泛收集和分析利用也带来了严重的隐私问题。实现数据价值开发与个人隐私保护之间的平衡，是现阶段学术界、产业界和政策领域着力解决的重大问题。目前广受关注的差分隐私保护方法主要通过扰动数据来避免因数据分析导致的隐私泄露。现有的差分隐私研究多假设存在一个“可信的”数据收集者进行数据处理。鉴于该假设在现实中较难成立，本项目研究先在终端扰动数据、再由收集者进行分析的“终端差分隐私保护模型”。本项目一方面探索影响个人隐私保护需求的各种要素，并定性定量分析隐私保护程度、分析结果的准确性与数据扰动程度之间的关系，研究扰动策略的优化方法；另一方面针对相似度计算等分析任务，研究可应对不同个体隐私需求差异的数据处理手段，探讨提高分析结果准确性的分析处理方法。本项目旨在提出实用有效的终端数据扰动方法和融合分析方法，为差分隐私保护的实用化提供理论、方法和技术支撑。

基于个人信息的大数据应用可创造巨大的价值，但个人数据的广泛收集和分析利用也带来了严重的隐私问题。实现数据价值开发与个人隐私保护之间的平衡，是现阶段学术界和产业界十分关注的问题。差分隐私是目前主流的隐私保护技术。传统的差分隐私研究多假设存在一个“可信的”数据收集者。鉴于该假设在现实中较难成立，本项目研究先在终端扰动数据、再由收集者进行分析的“终端差分隐私保护模型”。具体地，针对现有研究难以平衡隐私保护程度与数据可用性、难以满足个性化隐私保护需求、难以支撑复杂分析任务等问题，本项目从隐私预算的设置依据和方法、满足个性化隐私需求的终端差分隐私机制、面向相似度计算的终端差分隐私机制等几方面展开了研究。在隐私预算设置方面，本项目主要以基于终端差分隐私的推荐系统作为研究对象，将隐私预算的优化设置建模为多臂赌博机问题，并提出了基于置信度上界的学习策略，所提策略考虑到了用户对不同数据敏感程度不同的情况，可帮助用户在与服务器交互的过程中选择最优的隐私预算；在个性化隐私保护方面，本项目分别针对直方图估计和联邦学习这两类典型的数据分析场景，以隐私预算作为个人隐私需求的表征，建立了可满足个性化隐私需求的终端差分隐私模型，并对相应的数据聚合分析方法进行了改进，从而维持较高的准确率；在相似度计算方面，本项目主要针对协同过滤推荐、连续兴趣点推荐提出了满足终端差分隐私的推荐模型，针对特定的相似度计算方法和用户评分数据的特点对经典的扰动机制进行了改进，提出了可维持数据高可用性的评分扰动方法，针对签到数据的特点提出了可满足用户个性化隐私需求的数据扰动方法和分布估计方法。除了理论分析，本项目也开展了若干实验对所提方法的可行性和有效性进行了验证。上述研究成果有助于提升终端差分隐私保护方法的实用性，可为开发符合隐私保护相关法律法规的大数据应用提供理论支撑。