Mobility-Aware Multi-Model Fully Decentralized Federated Learning Framework for Vehicular Edge Computing

Abstract

Decentralized federated learning (DFL) has been introduced in vehicular networks to enable collaborative model training among clients through model exchange instead of sharing sensitive data, thereby preserving privacy and eliminating reliance on centralized infrastructure. Meanwhile, with the rapid growth of vehicular edge computing, an increasing number of deep learning models are deployed within a single vehicle to support diverse applications, leading to a multi-model learning setting. However, jointly training multiple models under dynamic mobility conditions is constrained by limited computational and communication resources, posing a significant challenge. To address these issues, this paper proposes a $K$ -constrained mobility-aware multi-model fully decentralized federated learning (KM-MMDFL) framework. The framework limits communication through a $K$ -communication constraint and incorporates a joint peer and model exchange selection strategy to reduce communication overhead while ensuring reliable and beneficial model exchange based on model quality and predicted connectivity. In addition, a utility-based model training mechanism is developed to improve training efficiency and fairness across task-specific models under limited resources. Experimental results in urban scenarios demonstrate that the proposed framework improves overall model accuracy across all task-specific models by up to 9.7 percentage points (16.3% higher) while reducing communication overhead by up to 65.4% compared to baseline methods.