Aggregation and perturbation in practice: Case-study of privacy, accuracy & performance

Abstract

We analyse accuracy, privacy, compression-ratio and computational overhead of selected aggregation and perturbation methods in the Internet of Things (IoT). We measure over a real-life data set of detailed energy consumption logs of a single family household. We modelled privacy by simple, threshold-driven machine-learning algorithms that extract features of behaviour. The accuracy of those extraction is used as privacy metric. We state for different parameters of the aggregation, reduction and perturbation if the output still allows detections, as this follows the EU's data protection principle of “minimisation”: increased privacy due to less detailed data, but still good enough accuracy for the purpose. The result is that many detections for sensible predictions and intelligent reactions are still possible with lower quality data.