Budget-Conscious Differentially Private Aggregation of Power Data Timeseries

Power consumption data collected in smart grids helps understand consumption trends and take informed decisions, such as how to allocate grid resources (e.g., transformers, storage units). However, electricity consumption data also discloses sensitive details about individuals, such as the times they are home, the kind of appliances they own, etc. Differential privacy (DP) is a protection model that adds noise to the data in a way that prevents an adversary from determining whether any specific individual has been included as part of the release or not. This process, called sanitization, has an inherent effect of reducing data accuracy as a trade-off for protection. For time series, maintaining reasonable levels of data accuracy becomes challenging, as multiple releases must be performed over time, and an attacker can correlate information from distinct timestamps to attack the privacy of an individual. We propose a novel approach that uses the Sparse Vector Technique (SVT) to judiciously allocate the amount of privacy budget available. Our approach brings two important advantages: a data analyst can obtain better accuracy compared to benchmarks for the same period of data release, or alternatively, the reporting period can be extended with a similar degree of accuracy. Extensive experiments on real data show that our proposed technique outperforms existing benchmarks.