Utility-Aware Time Series Data Release With Anomalies Under TLDP

With the prevalence of mobile computing, mobile devices have been generating numerous sensor data, a.k.a., time series. Since these time series may include sensitive information, users are posed with severe privacy risks. To protect individuals' privacy, local differential privacy (LDP) is proposed. However, the added noise satisfying LDP typically degrades the utility of released data, especially for anomaly detection such as healthcare monitoring and hazard alarming. In this paper, we study privacy-preserving time series release with anomalies. Recently, local differential privacy in the temporal setting (TLDP) is proposed to perturb the temporal order rather than the values. While it improves the utility for releasing value-critical data, it still suffers from low utility for anomaly detection, because of the inevitable missing and delayed values incurred by TLDP perturbation. We propose to improve its utility from two aspects. To reduce the missing values, we utilize selective substitution according to items' anomaly scores. To decrease the delayed values, we define metric-based $(\alpha , \delta )$ -TLDP and propose a mechanism that can prioritize anomaly release at a close timestamp while still guaranteeing the same TLDP privacy. Through theoretical and empirical evaluation, we show superior performance gain over existing TLDP-based mechanisms on both synthetic and real-world datasets.