Practical Crowdsourcing of Wearable IoT Data with Local Differential Privacy

In this work, we present and evaluate a crowdsourcing platform to collect wearable IoT data with local differential privacy (LDP). LDP protects privacy by perturbing data with noise, which may hinder their utility in some cases. For this reason, most researchers are wary of adopting it in their studies. To address these concerns, we consider the impact of different privacy budget values on the real wearable IoT data (steps, calories, distance, etc.) from N = 71 Fitbit users. Our goal is to demonstrate that, even if the collected information is protected with LDP, it is possible for data analysts to extract statistically significant insights on the studied population. To this end, we evaluate the error for various metrics of interest, such as sample average and empirical distribution. Furthermore, we verify that, in most cases, statistical tests produce the same results regardless of whether LDP has been applied or not. Our findings suggest that LDP with a privacy budget between 4 and 8 maintains an acceptable error of and over agreement on t-tests. Finally, we show that such values of privacy budget, albeit providing loose theoretical guarantees, can effectively defend against re-identification attacks on wearable IoT data.