Key-value data collection and statistical analysis with local differential privacy

Abstract

The collection and statistical analysis of simple data types (e.g., categorical, numerical and multi-dimensional data) under local differential privacy has been widely studied. Recently, researchers have focused on the collection of the key-value data, which is one of the main types of NoSQL data model. In the collection and statistical analysis of key-value data under local differential privacy, the frequency and mean of each key must be estimated simultaneously. However, achieving a good utility-privacy tradeoff is difficult, because key-value data has inherent correlation, and some users may have different numbers of key-value pairs. In this paper, we propose an efficient sampling based scheme for collecting and analyzing key-value data. Note that the more valid data collected, the higher the accuracy of statistical data under the same disturbance level and disturbance algorithm. Therefore, we make full use of probability sampling and the inherent correlation of key-value data to improve the probability of users submitting valid key-value data. Moreover, we optimize the budget allocation on key-value data, so that the overall variance of frequency and mean estimation is close to optimal. Detailed theoretical analysis and experimental results show that the proposed scheme is superior to existing schemes in accuracy.