Demo: A Practical Testbed for Decentralized Federated Learning on Physical Edge Devices
Chao Feng, Nicolas Huber, Alberto Huertas Celdran, Gerome Bovet, Burkhard Stiller
TL;DR
This paper tackles the challenge of realistically evaluating decentralized Federated Learning (DFL) on resource-constrained edge devices by introducing a physical, energy-aware testbed built on the NEBULA platform. The setup uses heterogeneous hardware (Raspberry Pi 4s and an NVIDIA Jetson Nano) with REST-based configuration, real-time metrics, and JT-TC66C energy meters to study DFL performance under multiple network topologies. Experiments on MNIST and FashionMNIST with a simple MLP trained via FedAvg over 10 rounds reveal that denser topologies yield higher $F_{1}$-scores (around 81–82%), while energy consumption remains within practical bounds; results on physical hardware are aligned with a virtualized baseline in accuracy but show slower training due to hardware constraints. The proposed testbed provides a realistic, energy-aware environment for evaluating DFL, enabling scalable topology studies and practical assessments of sustainability in edge-based federated learning.
Abstract
Federated Learning (FL) enables collaborative model training without sharing raw data, preserving participant privacy. Decentralized FL (DFL) eliminates reliance on a central server, mitigating the single point of failure inherent in the traditional FL paradigm, while introducing deployment challenges on resource-constrained devices. To evaluate real-world applicability, this work designs and deploys a physical testbed using edge devices such as Raspberry Pi and Jetson Nano. The testbed is built upon a DFL training platform, NEBULA, and extends it with a power monitoring module to measure energy consumption during training. Experiments across multiple datasets show that model performance is influenced by the communication topology, with denser topologies leading to better outcomes in DFL settings.