Randori: Local Differential Privacy for All

Abstract

Polls are a common way of collecting data, including product reviews and feedback forms. However, few data collectors give upfront privacy guarantees. Additionally, when privacy guarantees are given upfront, they are often vague claims about 'anonymity'. Instead, we propose giving quantifiable privacy guarantees through the statistical notion of differential privacy. Nevertheless, privacy does not come for free. At the heart of differential privacy lies an inherent trade-off between accuracy and privacy that needs to be balanced. Thus, it is vital to properly adjust the accuracy-privacy trade-off before setting out to collect data.Motivated by the lack of tools to gather poll data under differential privacy, we set out to engineer our own tool. Specifically, to make local differential privacy accessible for all, in this systems paper we present Randori, a set of novel open source tools for differentially private poll data collection. Randori is intended to help data analysts keep their focus on what data their poll is collecting, as opposed to how they should collect it. Our tools also allow the data analysts to analytically predict the accuracy of their poll. Furthermore, we show that differential privacy alone is not enough to achieve end-to-end privacy in a server-client setting. Consequently, we also investigate and mitigate implicit data leaks in Randori.

Efficient Error Prediction for Differentially Private Algorithms

“…As such, we believe the many variables' effect on error in our particular use case is difficult to investigate using methods from the current literature. Accordingly, we use Randori [13] as a use case where we create a prediction model for error. Randori is a set of tools for gathering poll data under local differential privacy [23].…”