Section: New Results

Protecting motion sensor data against sensitive inferences through an adversarial network approach

Participants : Antoine Boutet, Théo Jourdan.

With the widespread development of the quantified self movement, more and more motion sensor data are captured and transmitted through the intermediary of smartphones. However, granting to applications a direct access to sensor data expose users to many privacy risks, including in particular the possibility of inferring their activities and transportation mode to more sensitive inferences such as their demographic attributes or even mobility deficiency. In this work, we propose a privacy-preserving scheme to protect sensor data for activity recognition while at the same time preventing unwanted sensitive inferences on specific information. To achieve this objective, we leverage on the powerful framework of generative adversarial networks (GANs) to sanitize the sensor data. More precisely in our framework three neural networks are jointly trained, a generator that aim at sanitizing the data given at input as well two discriminators that try to infer respectively the sensitive attributes and the current activity of the user. By letting these neural networks compete against each other, the mechanism improves the protection while providing a good accuracy in terms of activity recognition and limiting sensitive inferences on specified attributes. Preliminary results demonstrate that the approach is promising in terms of achieving a good utility-privacy trade-off.