TC: Small: Adaptive Differentially Private Data Release

当前的信息技术使许多组织能够收集、存储和使用关于个人的大量和各种类型的信息。 各国政府和其他组织日益认识到分享如此丰富的信息的重要价值和巨大机会。 然而，数据隐私一直是这种信息共享的主要障碍，引起了人们对隐私保护数据发布和分析技术的关注。 差分隐私被广泛认为是最强的无条件隐私保证之一。虽然针对具有差异隐私的交互模型已经提出了许多有效的机制，但是具有差异隐私的非交互数据发布仍然是一个开放的问题，近年来只看到负面的结果。该项目旨在建立一个数据驱动和自适应的框架，用于差异化的私有数据发布。 它规避了在非交互式设置中的差异私有数据发布的难度，通过新颖和复杂的使用利用底层数据的特性的差异私有原语。 具体的研究目标包括：（1）设计自适应查询策略以发布具有差异隐私的数据，包括传统的关系数据和高维稀疏集值数据;（2）设计统计推断技术以使用发布的数据准确地回答用户查询;（3）设计算法以建模并将潜在的动态工作负载特性纳入框架中。 除了正式的分析和实验评估外，该项目还将在埃默里大学的真实的健康应用中评估和整合所开发的解决方案，以支持健康研究，同时提供严格的隐私保障。拟议研究的成功将有助于克服大规模数据共享的障碍，并将对数据隐私和信息管理领域以外的大型社会产生更广泛的影响。 该提案还包括一系列紧密结合的教育活动，包括关于数据隐私和安全的新课程开发，强调强有力的跨学科方面，继续让本科生参与研究，并鼓励妇女和少数民族参与。该项目还与埃默里大学在预测健康方面的战略举措密切相关，并将有助于开发新的博士学位。埃默里大学计算机科学与信息学专业。

Current information technology enables many organizations to collect, store, and use massive amount and various types of information about individuals. Governments and other organizations increasingly recognize the critical value and enormous opportunities in sharing such a wealth of information. However data privacy has been a major barrier for such information sharing, bringing much attention to privacy preserving data publishing and analysis techniques. Differential privacy is widely accepted as one of the strongest unconditional privacy guarantees. While many effective mechanisms have been proposed for the interactive model with differential privacy, non-interactive data release with differential privacy remains an open problem with the recent years only see negative results. This project aims to build a data-driven and adaptive framework for differentially private data release. It circumvents the hardness of differentially private data release in the non-interactive setting by novel and sophisticated use of the differentially private primitives exploiting the characteristics of the underlying data. The specific research objectives include: (1) design adaptive query strategies for releasing data with differential privacy, including traditional relational data and high dimensional and sparse set-valued data, (2) design statistical inference technique to accurately answer user queries using the released data, and (3) design algorithms to model and incorporate potentially dynamic workload characteristics in the framework. In addition to formal analysis and experimental evaluations, the project will evaluate and integrate the developed solutions in real health applications at Emory University to support health research while providing rigorous privacy guarantee.Success of the proposed research will help overcoming barriers for large scale data sharing and will have broader impacts to large societies beyond the field of data privacy and information management. The proposal also includes a set of closely integrated educational activities including new course development on data privacy and security emphasizing a strong interdisciplinary aspect, continued involvement of undergraduate students in research, and encouragement of women and minority for participation. The project also closely aligns with Emory's university-wide strategic initiatives in Predictive Health and will help develop the new Ph.D. program in Computer Science and Informatics at Emory University.