Pruning-Assisted Vehicular Federated Learning for CNN Training under Mobility-Aware Constraints

Abstract

Vehicular federated learning (VFL) has been widely adopted for collaborative edge training in vehicular networks to exploit massive onboard data while preserving data privacy and security. The dynamic nature of VFL training results in inconsistent client participation and unreliable timely model uploads due to vehicular mobility and system heterogeneity, which significantly degrades training efficiency and convergence performance. To address this issue, we propose a pruning-assisted VFL framework for convolutional neural network (CNN) training under mobility-aware uplink constraints, which leverages structured stochastic pruning to adaptively reduce both local computation complexity and uplink communication overhead. In this framework, we derive a per-round upper bound on the global loss function that explicitly characterizes the impact of the pruning-induced aggregation factor and the number of successfully uploading vehicles. Furthermore, the derived loss upper bound is minimized via a carefully designed multi-objective optimization formulation that jointly accounts for convergence contraction and stochastic noise effects. The resulting problem is formulated as a non-convex mixed-integer nonlinear program (MINLP) and is efficiently solved via an iterative algorithm based on continuous relaxation, equivalent reformulation, and successive convex approximation (SCA). Simulation results demonstrate that the proposed scheme substantially improves the successful upload probability and accelerates convergence of VFL, thereby achieving superior training efficiency compared to existing baseline methods.