Green Offloading and Model Size Adaptation for Federated Learning in IoT and Next-Generation Networks

Abstract

Federated learning (FL) trains models across geographically dispersed clients whose electricity carries very different carbon intensities. We present FedGreenCS, a carbon-aware approach to FL for IoT and next-generation networks. Clients are first grouped by data distribution similarity, using pairwise cosine similarity of model weights after a single local epoch. In each subsequent round, a sampled client whose carbon intensity exceeds a fixed tolerance offloads its training to the lowest-carbon available peer in its similarity cluster, while clients within the tolerance participate unchanged. Ordered dropout gives clients with a higher carbon intensity a narrower model. We derive a carbon cost model that separates computation and communication for each client and round, isolates the cost of the mandatory first-round broadcast, and states the design over the scaling rates and the tolerance. On a non-IID EMNIST partition with real regional carbon intensities, FedGreenCS with three carbon clusters reaches 70% test accuracy at 3.15 to 3.36 g CO2 of client-side carbon cost over three seeds, against 3.58 g for FedGreen and 15.12 g for FedAvg. With one model width for all clients, the offloading cuts this cost by 23% to 58%, but it raises the cost with two carbon clusters and always needs more rounds.