A Federated Learning-Based Edge Caching Approach for Mobile Edge Computing-Enabled Intelligent Connected Vehicles
A Federated Learning-Based Edge Caching Approach for Mobile Edge Computing-Enabled Intelligent Connected Vehicles
复制标题
适用于支持移动边缘计算的智能网联汽车的基于联合学习的边缘缓存方法
DOI：
10.1109/tits.2022.3224395
复制
发表时间：
2022-11-29
影响因子：
8.5
通讯作者：
Luo, Youlong
中科院分区：
文献类型：
--
作者：
Li, Chunlin;Zhang, Yong;Luo, Youlong
海量的地图数据传输和对高精度地图隐私的严格要求，给智能互联汽车(ICV)中的高精度地图缓存带来了重大挑战。为了减轻边缘网络的压力，保护隐私，引入了联邦学习(FL)。但汽车的高动力性和有限的资源导致精度低、训练延迟大。提出了一种联合学习中参与者选择和资源分配的联合优化方案。在每个时间片中，确定车辆是否参与训练，从而以有限的能量消耗最小化长期训练延迟。针对高精度地图缓存的时延和隐私要求，提出了一种基于联邦深度强化学习的边缘协作缓存方案，旨在实现动态自适应边缘缓存的同时保护用户隐私。协作缓存模型被描述为马尔可夫决策过程(MDP)。采用DUELING深度Q网络(DQN)求解最优策略，并利用联邦学习进行训练。有足够的对比实验来评估所提出的方案的性能。从可靠性、缓存命中率和训练精度等方面证明，该方法在满足高精度地图缓存的时延和可靠性要求的同时，有效地改善了联邦学习的训练参数。
Massive map data transmission and the strict demand for the privacy of high-precision maps have brought significant challenges to the cache of high-precision maps in intelligent connected vehicles (ICV). Federal learning (FL) was introduced to reduce the pressure on the edge network and protect privacy. But the high dynamics of cars and limited resources lead to low accuracy and high training delay. We propose a joint optimization scheme of participant selection and resource allocation for federated learning. In each time slice, vehicles are determined whether to participate in training, which minimizes long-term training delay with limited energy consumption. To meet the delay and privacy requirements of high-precision map caching, we present an edge cooperative caching scheme based on federated deep reinforcement learning (F-DRL), which aims to achieve dynamic adaptive edge caching while protecting user privacy. The collaborative caching model is formulated as a Markov decision process (MDP). Dueling Deep Q Network (Dueling-DQN) is used to solve the optimal strategy, and federal learning is used for training. Enough comparative experiments to evaluate the performance of the proposed schemes. The aspects of reliability, cache hit rate, and training accuracy prove that the method effectively improves the training parameters of federated learning while meeting a high-precision map cache's delay and reliability requirements.