Energy Efficiency Analysis of Federated Learning: Insights from UAV-Based Thermal Imaging Applications

Federated Learning (FL) is an approach for training a Machine Learning (ML) model with distributed training data and multiple participants. It allows bypassing the limitations of the traditional Centralized Machine Learning (CL) if data cannot be shared or stored centrally due to privacy restrictions or technical constraints - the participants train the model locally with their own training data and do not need to share it among the other participants. Only the locally updated model weights are shared after each round of training.
However, transitioning from CL to FL can inherently increase the demand for energy and computational resources. Additional resources are needed not only for the local training of each participant on its own device, but also for the communication and the server infrastructure to support and coordinate FL workflows.
This study investigates energy consumption in a real-world implementation of FL using different FL aggregation algorithms and workflows, using the NVIDIA Federated Learning Application Runtime Environment (NVFlare), compared to CL approach. It is also shown how the decentralized FL approaches (without a central server) behave compared to the centralized FL approaches, where a server is necessary. ... mehrFor this, a use case from the AI4EOSC project was used, involving uncrewed aerial vehicle (UAV)-based thermal images for detecting common thermal features in urban environments. The analysis leverages the Perun package to evaluate energy efficiency comprehensively across various FL workflows and aggregation algorithms by measuring their consumption. It is further explored how different aggregation strategies impact overall energy use and identifies the most energy-efficient FL methods under different operational conditions. Furthermore, each algorithm and workflow is evaluated to identify if the relationship between energy efficiency and effectiveness is significant enough to justify additional costs/efforts of our application.