A Siamese Adversarial Anonymizer for Data Minimization in Biometric Applications

Abstract

A wealth of sensors measure our day to day activities. Sharing these biometric data with cloud-enabled applications provides users with useful insights about their health status. However, due to their high re-identification potential, granting medical and fitness applications access to sensors data raises serious privacy concerns.

In this work, we aim to find a data representation that suppresses identity while retaining utility, i.e. data minimization. To this end, we define a component that is plugged into a machine learning pipeline and trained following a min-max adversarial optimization strategy. By using our component during training, we limit the need to rely on third parties and, therefore, reduce the risks of leakages and unintended processing. We evaluate privacy vs. utility trade-off on a real-world sharing scenario involving motion data, i.e. accelerometer and gyroscope measurements. Privacy is estimated by analyzing if an honest-but-curious attacker can effectively link two traces to the same user. In our experiments, we observe a decrease in test verification accuracy, i.e. the identifiability of the users, from 85% to 57%. This leaves an attacker with a representation that he can match only slightly better compared to a random guess. On top of that, the privatized gait representation is accompanied by an increase in performance w.r.t. the main tasks.