|
Abstract:
|
Federated Learning (FL) is an emerging area within Machine Learning (ML) that trains complex ML models using data that resides on a heterogeneous collection, or federation, of computing devices. The training is done in such a way that actual data never leaves the device that originally had it. In view of increasing privacy concerns and growing legal protections for digital privacy, interest in FL has been steadily increasing for the past several years. In this talk, I will discuss what Federated Learning can do for Mobile Health, and more broadly, for Digital Health. Furthermore, as privacy protection technologies and sophisticated privacy attacks both continue to proliferate, the question of how to balance algorithmic utility against privacy protection in a proactive manner becomes increasingly important. Between the extremes of no privacy protection and technologies that offer strong statistical guarantees (such as DP) at a high utility cost, we investigate a collection of "middle ground" techniques that provide strong protection against privacy threats specific to mHealth systems (especially systems employing FL) while maintaining reasonable algorithmic performance.
|
