Time series correlated key–value data collection with local differential privacy

Abstract

The data generated by users in various scenarios, such as video-sharing applications or smart home energy systems, requires robust privacy protection due to its sensitive nature. This includes estimating user behaviour over time, such as the proportion of users watching video, the average watching ratio, or household energy consumption and average electricity usage. After privacy protection is applied, the processed data is used to analyse user behaviour and optimize systems. However, this specific requirement for high accuracy in frequency and mean estimation after privacy protection is not effectively addressed by existing methods. To fill this gap, the Time Correlated Key–Value with Local Differential Privacy (TSCKV) is proposed in this paper. A tighter privacy budget composition bound is obtained by a perturbation scheme that exploits key–value ($k-v$) pair correlations while sacrificing some of the value data. By setting a threshold, values that change below it can be set to zero directly, saving the privacy budget. Estimators and correctors for the $k-v$ pairs are proposed by this work. Using the real Kuairec dataset, experiments show that the overall statistical utility of TSCKV, including frequency and mean estimation, is higher than that of the time series data mechanism alone and the $k-v$ pair mechanism with simple privacy budget allocation. Additionally, TSCKV achieves more accurate early frequency estimation compared to the static $k-v$ pair correlated perturbation mechanism.