Computing Statistics from Private Data

Abstract

In several domains, privacy presents a significant obstacle to scientific and analytic research, and limits the economic, social, health and scholastic benefits that could be derived from such research. These concerns stem from the need for privacy about personally identifiable information (PII), commercial intellectual property, and other types of information. For example, businesses, researchers, and policymakers may benefit by analyzing aggregate information about markets, but individual companies may not be willing to reveal information about risks, strategies, and weaknesses that could be exploited by competitors. Extracting valuable utility from the new "big data" economy demands new privacy technologies to overcome barriers that impede sensitive data from being aggregated and analyzed.Secure multiparty computation (MPC) is a collection of cryptographic technologies that can be used to effectively cope with some of these obstacles, and provide a new means of allowing researchers to coordinate and analyze sensitive data collections, obviating the need for dataowners to share the underlying data sets with other researchers or with each other. This paper outlines the findings that were made during interdisciplinary workshops that examined potential applications of MPC to data in the social and health sciences.The primary goals of this work are to describe the computational needs of these disciplines and to develop a specific roadmap for selecting efficient algorithms and protocols that can be used as a starting point for interdisciplinary projects between cryptographers and data scientists.

It is not (only) about privacy: How multi-party computation redefines control, trust, and risk in data sharing

“…Moreover, unlike MPC, which uses encryption, differential privacy protects the data by adding random noise during the analysis. Nevertheless, those technologies can complement each other to implement robust security requirements in various use cases (e.g., Alter et al, 2018;Pettai & Laud, 2015;Zhong et al, 2020).…”

Information-theoretic metrics for Local Differential Privacy protocols

“…[4,6,31,45] define privacy metrics based on min-entropy, representing a worst-case privacy loss, and bound this from above by ε-DP. 2 Continuing in this line, [3] uses a min-entropy utility measure and gives an upper bound on utility in terms of ε-DP. Rényi-DP unifies entropic and min-entropic privacy [42].…”

“…Secure Multiparty Computation (SMC). Protocols for tallying secret values are well studied in the field of SMC [48,38,2]; they are not computationally expensive but require more communication than LDP. While LDP protocols always have nonzero leakage, SMC can offer strong privacy guarantees in many situations.…”