arXiv paper: GQ-FSL: Green Quantized Federated Split Learning

Researchers Idan Roth and Lutz Lampe propose Green Quantized Federated Split Learning (GQ-FSL), a framework that combines federated split learning with stochastic quantization to cut energy use on resource-constrained mobile devices while maintaining accuracy. It allows different numerical precisions for client- and server-side submodels—“asymmetric precision”—so device energy limits can be managed without proportionally hurting global model quality. The authors supply parameterized energy models for the split architecture and a theoretical convergence bound under non-identically distributed data. They frame the choice of split point and precision levels as a joint optimization that minimizes total system energy consumption subject to a target accuracy. In their study, GQ-FSL shows better energy efficiency than quantized federated learning or full-precision federated split learning, making it a potential enabler for deploying large deep neural networks at the wireless edge under strict energy budgets. The work is to appear at the 2026 IEEE SPAWC workshop.