To Share or Not to Share: On Location Privacy in IoT Sensor Data

Abstract

Data sharing is crucial for building large datasets which in return are essential for developing and training accurate models in many contexts including smart cities, agriculture, and medical applications. However, shared data may leak private information, such as personal identifiers or location. Past research provides evidence that solely removing these identifiers through pseudonymization is not enough to ensure data privacy protection, since even the pseudonymized data may still contain information about the data provider. In this paper, we show that sensor data may leak a sensor's location even if the latter is not explicitly shared. Sensors are localized by linking sensor data with publicly available environmental data such as local weather. The proposed localization method relies on a machine learning model to predict weather data from sensor observations. Subsequently, the localization algorithm determines the sensor's location from the predicted weather trace using Bayesian filtering. We apply our approach to three real-world datasets where we (1) localize an ozone sensor given its readings, (2) localize a cow from activity parameters recorded with a tracker in the cow's reticulum, (3) localize solar panels based on their solar generation data. The achieved average localization accuracy of 5.68 km, 19.91 km, and 13.68 km on the above tasks, respectively, using data traces with a length of 365 days is remarkable. In addition, we introduce a mechanism, referred to as teleport, to protect location information in sensor data. The mechanism is based on deep models and masks the location by replacing the weather dependency with a different weather signature.