Hermes: A Privacy-Preserving Approximate Search Framework for Big Data

Abstract

We propose a sampling-based framework for privacy-preserving approximate data search in the context of big data. The framework is designed to bridge multi-target query needs from users and the data platform, including required query accuracy, timeliness, and query privacy constraints. A novel privacy metric, $(\varepsilon,\delta)$ -approximation, is presented to uniformly measure accuracy, efficiency and privacy breach risk. Based on this, we employ bootstrapping to efficiently produce approximate results that meet the preset query requirements. Moreover, we propose a quick response mechanism to deal with homogeneous queries, and discuss the reusage of results when appending data. Theoretical analyses and experimental results demonstrate that the framework is capable of effectively fulfilling multi-target query requirements with high efficiency and accuracy.