Private Data Exploration with Provable Guarantees

计算机技术使我们个人生活的大量数字痕迹得以收集和存储。许多研究项目和现实生活中的应用至关重要地依赖于对这些数据集的探索，但最近欧洲通过的《通用数据保护条例》（GDPR）和有争议的Facebook隐私丑闻表明，隐私问题非常严重。一个简单的临时隐私保护方法可能会给个人和公司带来巨大的成本。现有的数据库系统具有强大的隐私保证（如差分隐私），但对分析人员真正关心的问题（包括查询答案中的错误（准确性）和收到答案之前的等待时间（延迟））没有提供直接的、最优的保证。由于缺乏隐私背景，分析人员在使用这些系统时很容易遭受准确性差或等待时间长的问题。将现有的数据探索技术（如抽样或在线聚合）天真地应用于私有设置，要么会破坏隐私保证，要么会导致准确性低下。而且，这些系统提供的隐私保障是相同的，不能满足不同应用的不同隐私需求。

Computing technology has enabled massive digital traces of our personal lives to be collected and stored. Numerous research projects and real-life applications crucially rely on the exploration of these datasets, but there are great privacy concerns as demonstrated by the recent passing of the General Data Protection Regulation (GDPR) in Europe and the controversial Facebook privacy scandal. A simple ad-hoc privacy-preserving approach can lead to a drastic cost to individuals and companies. Existing database systems with strong privacy guarantee such as differential privacy provide no direct, optimal guarantees for what the analysts really care, including the error in the query answer (accuracy) and the waiting time before receiving the answer (latency). With insufficient privacy background, the analysts can easily suffer poor accuracy or long waiting time when using these systems. Naively applying existing data exploration techniques, such as sampling or online aggregation, to the private setting can either break the privacy guarantee or result in poor accuracy. Moreover, these systems offer the same privacy guarantee and cannot meet the diverse privacy requirements of different applications. My proposed research aims primarily at enabling private data exploration with provable guarantees such that all parties involved are not required to be privacy experts -- data analysts can naturally explore sensitive data with a rich set of query templates and functionalities while data curator can easily track the privacy loss over the entire exploration process. This research program will build on my current work on customized privacy and private exploration systems. To do so, my research program includes a set of short-term and long-term research tasks. The short-term research plan aims to (1) develop accuracy-aware data explorations under the state-of-the-art privacy guarantee, differential privacy, and (2) enable private sampling and online aggregation to enhance the interactive user experience during data exploration. The long-term research plan is to support customizable privacy guarantees other than differential privacy in private data exploration. The proposed frameworks will be tested through prototyping with open-sourced database systems.