CIF: Small: Privacy and Utility of Databases: An Information-Theoretic Approach

信息技术和电子通信已迅速应用于人类活动的各个领域，包括商业、医药和社交网络。伴随而来的无数大型集中式可搜索数据存储库的出现，使得通过数据关联（无意或恶意设计）导致的私人信息“泄露”成为一个重要而紧迫的社会问题。在提供必要的隐私保证的同时保持这些数据源的有用性是一个尚未解决的重要问题。这个问题促使我们需要一个全面的分析框架，它可以明确地告诉我们私有数据有多安全（隐私），同时仍然为多个合法的信息消费者提供有用的好处（效用）。本研究开发了一个统一的框架来研究效用-隐私权衡，而不考虑数据源类型或扰动方法。比率失真理论的技术和结果用于数据源建模，开发独立于应用程序的效用和隐私度量，并开发用于处理外部知识问题的侧信息模型。将适用于单查询数据源模型的框架扩展到研究多查询模型的实用-隐私权衡。还研究了一个连续披露问题，该问题利用连续细化中的经典结果来开发多个查询不会导致额外信息丢失的条件。开发的通用框架包括工具和技术，以弥合信息理论模型和当前方法与计算机科学中占主导地位的理论框架之间的差距。

Information technology and electronic communications have been rapidly applied to every sphere of human activity, including commerce, medicine and social networking. The concomitant emergence of myriad large centralized searchable data repositories has made "leakage" of private information via data correlation (inadvertently or by malicious design) an important and urgent societal problem. Maintaining the usefulness of these data sources while also providing necessary privacy guarantees is an important unsolved problem. This problem drives the need for an overarching analytic framework that can tell us unequivocally how safe private data can be (privacy) while still providing useful benefit (utility) to multiple legitimate information consumers.This research develops a unified framework to study the utility-privacy tradeoff irrespective of the type of data source or method of perturbation. Techniques and results from rate-distortion theory are used to model data sources, develop application independent utility and privacy metrics, and develop a side-information model for dealing with questions of external knowledge. The framework, applicable for single query data source models, is extended to study the utility-privacy tradeoffs for multiple-query models. Also studied is a successive disclosure problem which draws on classic results in successive refinement to develop the conditions under which multiple queries result in no additional information loss. The universal framework developed includes tools and techniques to bridge the gap between the information-theoretic model and current approaches and the dominant theoretical framework in computer science.