Assessing Energy Consumption and Thermal Impact in IoT Edge Federated Learning

Federated Learning (FL) enables distributed training of Machine Learning (ML) models directly on Edge Devices, preserving data privacy and reducing cloud dependency. This innovative approach is particularly valuable in Internet of Things (IoT)-aware pervasive environments, as Edge Computing enables low-latency and efficient decentralized learning. However, deploying FL on resource-constrained devices introduces several challenges related to energy consumption and thermal management. Many studies evaluate FL through simulations, often neglecting its actual hardware impact. This work proposes an experimental setup on real devices to fully assess the practical viability of Federated Learning on resource-constrained IoT edge devices. Specifically, we quantify the impact of FL on energy consumption, execution time, and thermal performance to provide recommendations for creating sustainable real-world Federated Learning settings. Our experimental setup utilises a portable USB meter and software monitoring tools to keep track of energy consumption and metrics such as execution time and device temperature. The Brain Tumor MRI dataset is used to implement a Healthcare domain FL scenario, where data privacy and energy consumption tracking are essential. This study aims to bridge the gap between theoretical FL simulations and real-world deployments by analyzing the impact of Federated Learning in IoT-edge devices. Training local ML models in FL scenarios relies on numerous mathematical operations leading to increased computational load and increased energy consumption. Our study identifies critical phases, such as idle periods between rounds and the final rounds as the model approaches convergence, that can be leveraged to implement adaptive thermal control mechanisms, and optimize energy consumption and contribute to sustainable IoT-driven FL.