A Survey on Randomized Mechanisms for Statistical Learning under Local Differential Privacy

Abstract

High confidence computing is an emerging computing paradigm, which supports the next-generation information system. Statistical learning is a basic research direction of this paradigm, which has been widely applied to various services and applications, such as data mining and data analysis. However, data privacy is one of the most concerned challenges by multiple participants and has an important influence on the development of statistical learning. Local differential privacy is a strong privacy metric in a local setting, which has a rigorous mathematical model. There are a series of research works to propose efficient randomized mechanisms for statistical learning. However, there are few literatures to provide a systematic overview of the existing works.In this paper, we present the existing literatures for randomized mechanisms under local differential privacy. At first, we introduce the basic perturbation mechanisms satisfying local differential privacy. Then, we give a framework and the main components of locally private mechanisms. Next, we describe the detailed content of randomized mechanisms for frequency estimation, heavy hitter identification and distribution estimation. Finally, we discuss the possible trends and challenges of the future research. Our survey provides a systematic and comprehensive understanding about randomized mechanisms for statistical learning under local differential privacy.