FedSynthesis: A Flower-based Framework for Carbon-Reduced Federated Learning

New machine learning paradigms, such as Federated Learning (FL), have become popular for processing privacy-sensitive information without leaking confidential data to unintended parties in networked settings. However, they introduce additional computation and communication overhead. Such overhead leads to an increase in carbon emissions, threatening the sustainability of systems and jeopardizing net-zero goals. This work proposes a carbon-aware FL framework by extending the Flower framework via the incorporation of customized aggregation strategies for carbon emission reduction. For this purpose, an empirical approach is adopted, which leverages the impact of the carbon emission tracker tool CodeCarbon, allowing for the development of new methods based on measured carbon values. Furthermore, MLFlow, an MLOps tool, is integrated into the framework, enabling users to collect metrics and visualize them. The applicability of this framework is tested on a previously published security scheme and federates its machine learning pipeline. A comparison of this approach against a standard FL setup indicates improvements in carbon reduction.