Implementation of Federated Learning on Resource-constrained devices: Lessons learned

In this paper, we implement and deploy the most widely used algorithm in Federated Learning (FL), i.e. Federated Averaging (FedAvg), on an experimental testbed. The testbed consists of Raspberry Pi devices (RPis) connected to a wireless network. We perform extensive evaluations in various real-world scenarios and provide insightful results and lessons learned. Specifically, we evaluate FedAvg on our testbed to investigate the effect of the following parameters on training: (i) number of users selected at each round, (ii) number of local gradient steps before communicating with the server, (iii) clients disconnecting from the server, (iv) data distribution heterogeneity across clients, and (v) mobility of users. Finally, we examine the impact of the wireless environment on the learning performance under varying network parameters.