Eco-Friendly AI: a framework for Data Centric Green Federated Learning

The environmental cost of deep learning is increasingly significant, prompting a shift from performance-driven "Red AI" to sustainability-focused "Green AI". In this paper, we propose a data-centric framework for environmentally sustainable Federated Learning (FL), focused on optimising data quality and node selection to minimise carbon emissions without compromising model performance. Central to our approach is an interactive FL Configuration Selection System, which, given dataset and infrastructure characteristics, assists researchers in configuring greener FL training workflows. Our system integrates data quality metrics and carbon footprint estimates to select environmentally optimal nodes and applies intelligent data reduction through three strategies: Node Selection, Minimal Smart Reduction, and Smart Reduction. We demonstrate the effectiveness of our tool in the context of time series classification, offering a practical solution for sustainable FL research.