Efficient Federated Learning via Adaptive Model Pruning for Internet of Vehicles With a Constrained Latency

Abstract

In the Internet of Vehicles (IoV), data privacy concerns have prompted the adoption of Federated Learning (FL). Efficiency improvements in FL remain a focal area of research, with recent studies exploring model pruning to lessen both computation and communication overhead. However, in the IoV, model pruning presents unique challenges and remains underexplored. Pruning strategy design is critical as it directly impacts each vehicle's learning latency and capacity to participate in FL. Furthermore, FL performance and model pruning are intricately connected. Additionally, the fluctuating number and mobility states of vehicles per round complicate determining the optimal pruning ratio, closely intertwining pruning with vehicle selection. This study introduces Vehicular Federated Learning with Adaptive Model Pruning (VFed-AMP) to tackle these challenges by integrating adaptive pruning with dynamic vehicle selection and resource allocation. We analyze the impact of pruning ratios on learning latency and convergence rate. Then, guided by these findings, a joint optimization problem is formulated to maximize the convergence rate concerning optimal vehicle selection, bandwidth allocation, and pruning ratios. Finally, a low-complexity algorithm for joint adaptive pruning and vehicle scheduling is proposed to address this problem. Through theoretical analysis and system design, VFed-AMP enhances FL efficiency and scalability in the IoV, offering insights into optimizing FL performance through strategic model adjustments. Numerical results on various datasets show VFed-AMP achieves superior training accuracy (e.g., at least 13.4% improvement for BelgiumTS) and significantly reduces training time (e.g., at least up to $1.8\times$ for CIFAR-10) compared to traditional FL methods.