Fitness-Driven Evolutionary Federated Learning: Adaptive Client Selection and Dynamic Population for Communication Efficiency

Federated Learning (FL) enables decentralized nodes to collaboratively train models while maintaining data privacy. However, traditional FL methods often face significant challenges, including high communication overhead and slow convergence, especially when data across nodes is heterogeneous. To address these issues, this paper introduces a novel framework named Fitness-Driven Evolutionary Federated Learning (FD-EFL), which effectively combines evolutionary strategies (ES) with adaptive client selection and dynamic population adjustment. The primary innovation of FD-EFL is its fitness-driven information-sharing approach, wherein clients communicate only concise fitness metrics representing similarities between their local models and a noise-perturbed global population. This significantly reduces communication overhead. FD-EFL further enhances performance by adaptively selecting high-quality clients, effectively minimizing the impact of noisy or low-quality updates. Additionally, the framework integrates a dynamic population adjustment mechanism guided by the Critical Learning Period (CLP), dynamically expanding the population size during critical training phases to improve model accuracy, and shrinking it during non-critical phases to save communication resources. Experimental evaluations demonstrate that FD-EFL substantially lowers communication costs without compromising model accuracy, achieving comparable performance to established methods like FedAvg. Our proposed framework offers a practical and efficient solution for federated learning in heterogeneous data environments. The implementation is publicly available at: https://github.com/buaaYYC/Fed-FEL.