Open Challenges in Federated Machine Learning

Abstract

Federated machine learning is an innovative technique to allow one to train machine learning models mainly on distributed (user) devices not to share private data with third parties. Each device contributes by training a partial model based on local data; a centralized orchestrator aggregates the results. Federated machine learning enhances data privacy and reduces network overhead, but it requires complex coordination mechanisms that handle many devices connected to a potentially unstable network. While this approach is gaining more and more traction, this article tries to summarize and highlight the key challenges that are still open and calls for further contributions from both academia and industry.