Green Federated Learning: Quantifying The Energy-Accuracy Trade-Off In Decentralized Iot Networks

Abstract

Artificial intelligence systems have become increasingly energy-intensive. Deep learning model training consumes substantial electrical power across the global computing infrastructure. The environmental impact of such computational demands raises serious sustainability concerns within the machine learning community. Federated learning enables decentralized model training across distributed edge devices. Local computations occur on individual nodes without centralizing raw data. Privacy preservation and reduced communication overhead represent primary advantages of federated architectures. However, aggregate energy consumption across millions of participating edge devices remains poorly characterized. The Green-FL protocol introduces energy-awareness into federated learning optimization objectives. Dynamic resource allocation mechanisms classify devices based on real-time power source availability. Active nodes connected to renewable energy receive priority for training tasks. Battery-powered devices enter standby states until sustainable power becomes available. Training schedules adapt to fluctuations in clean energy availability across device networks. Experimental evaluation demonstrates substantial energy consumption reduction without significant accuracy degradation. Convergence time increases moderately due to intermittent node availability during low-renewal periods. Carbon emission projections indicate meaningful environmental benefits at the deployment scale. The accuracy-per-watt optimization metric provides a quantifiable framework for sustainable machine learning development. Edge computing environments benefit particularly from such energy-conscious training protocols.