Local Differential Privacy Protocol for Making Key–Value Data Robust Against Poisoning Attacks

Abstract

Local differential privacy is a technique for concealing a user’s information from collectors by randomizing the information within the user’s own device before sending it to unreliable collectors. Ye et al. proposed PrivKV, a local differential privacy protocol for securely collecting key–value data, which comprises two-dimensional data with discrete and continuous values. However, such data is vulnerable to a “poisoning attack,” whereby a fake user sends data to manipulate the key-value dataset. To address this issue, we propose an Expectation-Maximization (EM) based algorithm, in conjunction with a cryptographical protocol for ensuring secure random sampling. Our local differential privacy protocol, called emPrivKV, offers two main advantages. First, it is able to estimate statistical information more accurately from randomized data. Second, it is robust against manipulation attacks such as poisoning attacks, whereby malicious users manipulate a set of analysis results by sending altered information to the aggregator without being detected. In this paper, we report on the improvement in the accuracy of statistical value estimation and the strength of the robustness against poisoning attacks achieved by applying the proposed method to open datasets.