Privacy-Preserving Detection of Anomalous Phenomena in Crowdsourced Environmental Sensing

Abstract

Crowdsourced environmental sensing is made possible by the wide-spread availability of powerful mobile devices with a broad array of features, such as temperature, location, velocity, and acceleration sensors. Mobile users can contribute measured data for a variety of purposes, such as environmental monitoring, traffic analysis, or emergency response. One important application scenario is that of detecting anomalous phenomena, where sensed data is crucial to quickly acquire data about forest fires, environmental accidents or dangerous weather events. Such cases typically require the construction of a heatmap that captures the distribution of a certain parameter over a geospatial domain (e.g., temperature, ${\text {CO}}_{2}$ concentration, water polluting agents, etc.).

However, contributing data can leak sensitive private details about an individual, as an adversary may be able to infer the presence of a person in a certain location at a given time. In turn, such information may reveal information about an individual’s health, lifestyle choices, and may even impact the physical safety of a person. In this paper, we propose a technique for privacy-preserving detection of anomalous phenomena, where the privacy of the individuals participating in collaborative environmental sensing is protected according to the powerful semantic model of differential privacy. Our techniques allow accurate detection of phenomena, without an adversary being able to infer whether an individual provided input data in the sensing process or not. We build a differentially-private index structure that is carefully customized to address the specific needs of anomalous phenomenon detection, and we derive privacy-preserving query strategies that judiciously allocate the privacy budget to maintain high data accuracy. Extensive experimental results show that the proposed approach achieves high precision of identifying anomalies, and incurs low computational overhead.