Joint resource management for mobility supported federated learning in Internet of Vehicles
Joint resource management for mobility supported federated learning in Internet of Vehicles
复制标题
DOI：
10.1016/j.future.2021.11.020
复制
发表时间：
2022-04-01
影响因子：
7.5
通讯作者：
Zhao, Chenglin
中科院分区：
文献类型：
--
作者：
Wang, Ge;Xu, Fangmin;Zhao, Chenglin
近年来，多通道边缘计算(MEC)和人工智能(AI)的强大结合，即边缘智能，推动了智能交通系统(ITS)的发展。然而，在车载边缘场景下，日益增强的消费者隐私意识与集中式AI训练方案中的数据泄露风险之间存在不匹配，成为满足用户体验的新障碍。作为一种很有前景的隐私保护范型，联邦学习只需利用分散训练的局部模型的参数来合成全局模型，避免了敏感数据的暴露。鉴于此，我们将联邦学习引入到所提出的两级MEC辅助车载网络框架中。本文旨在解决将联合学习引入车联网(IoV)场景所带来的挑战。首先，车辆作为参与者的实体(联邦学习的局部模型训练节点)，具有较高的机动性。我们设计了一种支持移动性的联合学习参与者决策算法来从候选车辆中挑选参与者。其次，联合学习相当消耗资源，不可避免地会给参与者带来相当大的成本。重点研究了联合资源分配问题，以优化联合学习成本。最后，针对集中式资源分配的局限性，提出了一种受多智能体深度强化学习启发的全分布式资源分配方法。仿真结果验证了所提方案的可行性和有效性。(C)爱思唯尔出版的《2021年》。
In recent years, the powerful combination of Multi-access Edge Computing (MEC) and Artificial Intelligence (AI), called edge intelligence, promotes the development of Intelligent Transportation Systems (ITS). However, there is a mismatch between the ever-increasing consumer privacy awareness and the data leakage risk in centralized AI training solutions in vehicular edge scenarios, which has become a new obstacle to satisfying the user experience. As a promising privacy-preserving paradigm, federated learning synthesizes a global model only with the parameters of decentralized trained local models, avoiding the exposure of sensitive data. Given this, we introduce federated learning into the proposed two-level MEC-assisted vehicular network framework. This paper aims to address the challenges posed by adopting federated learning into the Internet of Vehicles (IoV) scenario. Firstly, as the entity of the participant (the local model training node of federated learning), vehicles have high mobility. We design a mobility supported federated learning participant decision algorithm to pick out participants from candidate vehicles. Secondly, federated learning is rather resource-consuming, inevitably incurring considerable costs to participants. We focus on the joint resource allocation problem to optimize the federated learning cost. Finally, considering the limitations of centralized resource allocation, we propose a fully distributed resource allocation method inspired by multiagent deep reinforcement learning. Simulation results are presented to demonstrate the feasibility and effectiveness of the proposed schemes. (C) 2021 Published by Elsevier B.V.