Network Slice-Enabled Federated Learning for Collaborative Traffic Perception in Vehicular Edge Computing

This paper presents a novel network slice-enabled federated learning framework for collaborative traffic perception in vehicular edge computing environments. The proposed approach integrates channel-aware federated learning with intelligent resource allocation to enhance perception performance across heterogeneous vehicular networks. We establish a comprehensive channel model incorporating Doppler effects, multipath fading, and temporal-spatial variations to characterize vehicular communication dynamics. The framework employs three specialized perception slices for visual environment perception, radar environment perception, and semantic perception, each optimized through end-to-end neural network architectures. A reinforcement learning-based resource allocation mechanism dynamically manages computational, memory, and energy resources while considering real-time channel conditions. Extensive simulations across three resource environments demonstrate that the proposed PPO-based approach achieves superior performance with 94% task completion rate, 32ms communication-computation latency, and 3.35 energy efficiency ratio in high-resource scenarios. The results validate the framework’s effectiveness in managing federated learning convergence and resource optimization under varying vehicular network conditions, significantly outperforming traditional SAC and heuristic PSO approaches.