Reducing The Carbon Footprint Of Federated Learning Through Intelligent Client Selection

Abstract

Federated learning (FL) spreads model training across distributed clients located in different geographical areas, each utilizing heterogeneous energy sources with varying carbon intensities. Uniform random client sampling or selection solely on data utility fails to consider the carbon footprint incurred in the training, which leads to unnecessarily high carbon emissions whenever a high-carbon client is selected again. In this paper, GreenFL—an intelligent client selection framework—optimally selects clients to achieve joint improvements in model accuracy, convergence speed, and carbon footprint by incorporating the real-time carbon intensity signal of regional grid power APIs into FL round scheduling. GreenFL proposes a novel Carbon-Utility Score to trade off data diversity contribution and per-client carbon intensity with the help of a dynamically fluctuating environmental pressure factor, which grows rapidly when aggregate emissions reach the predefined carbon budget. Tested on CIFAR-10 and FEMNIST among 200 simulated heterogeneous clients located in 5 carbon intensity regions, GreenFL brings a 62% emission reduction compared to FedAvg without impacting the accuracy significantly to 90.1%, which outperforms EcoLearn's 61% emission reduction to 88.7% accuracy.