SCPFL: A Safety Reinforcement Learning Framework for Joint Energy-Fairness Optimization in Long-term Federated Learning

Abstract

Federated Learning (FL) enables heterogeneous clients to collaboratively train a shared model while keeping raw data local. In practical FL systems, client participation is often constrained by limited battery energy, and long-term unselection of clients may lower their willingness to participate, which further reduces the number of training data. Existing studies mainly focus on either energy-aware training or fairness-aware participation, while ignoring to design a long-term FL from both aspects. To address this issue, this paper proposes a Safe Client Pool FL framework (SCPFL) to jointly optimize client selection and local iteration under both energy and fairness constraints. Specifically, we first model the long-term energy consumption and fairness participation as probabilistic constraints. Based on these two constraints, we further prove that the conditions to obtain feasible client sets by satisfying either energy or fairness constraint. Finally, the considered problem is converted to an unconstrained long-term optimization problem, which can be solved by designed Actor-Critic (AC) architecture. Experiments on MNIST, Fashion-MNIST, CIFAR-10, CIFAR-100, and UCF-101 show that SCPFL consistently outperforms baselines and improves the energy-training performance by 11% on average.