CAREER: An Integrated Approach For Efficient Privacy Preserving Distributed Data Analytics

越来越多的不同组织需要安全地共享他们的私有数据来执行许多关键任务。最近，基于安全多方计算（SMC）和数据清理技术的几种不同方法已经出现，以实现保护隐私的分布式数据分析。尽管基于SMC的隐私保护协议允许参与方只学习最终（准确）结果，但它们不能很好地扩展到大量数据。另一方面，基于清理的技术允许组织在某些隐私保证下通过扭曲数据来暴露隐私敏感数据。在许多情况下，保护隐私所需的严重数据失真可能导致不准确的结果。由于现有方法的局限性，需要高效、准确的隐私保护解决方案来处理大型分布式数据集。为了应对这一挑战，我们设计并开发了一个新的框架，将消毒和SMC技术集成在一起，以在资源限制下开发有效的隐私保护解决方案。基本上，我们使用数据消毒技术来获得初始近似结果，并有选择地进行SMC操作以提高准确性。由于我们以黑盒方式使用现有技术，因此我们的方法与任何新的消毒或SMC技术正交。我们的新技术将大大降低执行保护隐私的分布式数据分析协议的成本。这将通过为新的应用（例如电子保健和电子政务应用）开辟道路而产生直接的经济影响，这些应用目前由于缺乏能够在大型数据集上有效工作的必要的隐私保护解决方案而被认为是不可行的。

Increasingly, different organizations need to securely share their private data to execute many critical tasks. Recently, several different approaches based on secure multi-party computation (SMC) and data sanitization techniques have emerged to enable privacy preserving distributed data analytics. Although SMC based privacy-preserving protocols allow the participating parties to learn only the final (accurate) result, they do not scale well for large amounts of data. On the other hand, sanitization based techniques allow organizations to reveal privacy sensitive data under some privacy guarantees by distorting the data. In many cases, significant data distortion that is needed to preserve privacy could lead to inaccurate results. Due to the limitations of the current approaches, efficient and accurate privacy-preserving solutions are needed for handling large distributed data sets. To address this challenge, we design and develop a novel framework where sanitization and SMC techniques are integrated to develop efficient privacy-preserving solutions under resource constraints. Basically, we use the data sanitization techniques to get initial approximate results and carry out SMC operations selectively to increase the accuracy. Since we use existing techniques in a black box fashion, our approach is orthogonal to any new sanitization or SMC techniques.Our new techniques will substantially decrease the cost of executing privacy-preserving distributed data analytics protocols. This will have a direct economic impact by opening the way for new applications (e.g., e-health and e-government applications) that are at present considered infeasible due to the lack of necessary privacy-preserving solutions that can work efficiently on large data sets.