Analyzing Federated Learning Performance in Distributed Edge Scenarios

Abstract

Federated Learning is a machine learning paradigm where many clients cooperatively train a single centralized model while keeping their data private and decentralized. This novel paradigm imposes many challenges, such as dealing with data that is not independent and identically distributed, spread among multiple clients that are not synchronized and may have limited computing power. These clients are often edge devices such as smartphones and sensors, which form a system that is heterogeneous, highly distributed by nature and difficult to manage. This work proposes an architecture for running federated learning experiments in a distributed edge-like environment. Based on this architecture, a set of experiments are conducted to analyze how the overall system performance is affected by different configuration parameters and varied number of connected clients.