| In recent years,with the development of computing power and network communication technology,more and more distributed artificial intelligence applications are carried on smart terminals such as smartphones,wearable devices,and autonomous vehicles.Training often results in a steep rise in network communication overhead.To this end,this paper attempts to jointly train machine learning models without exchanging data by deploying edge-to-edge collaborative distributed machine learning in a wireless network environment,reducing communication overhead and service delay and protecting user data privacy.In response to the above requirements,this paper first proposes an edge-to-end collaborative distributed training framework based on hierarchical federated learning,then designs a strategy for generating aggregated communication network topology under local area networks,and finally implements a distributed training system for mobile terminals.The main work of this paper is as follows:(1)A side-end collaborative distributed training framework based on hierarchical federated learning is proposed.The terminal grouping method for information perception in the local area network is designed,which relieves the pressure of storage and communication in the data center,improves the security of data privacy,and reduces the cost of cross-WAN communication.(2)The generation strategy of aggregated communication network topology under local area network is proposed.The topology can be divided into centralized network topology and decentralized network topology.According to the real-time network information in the local area network and the computing power information of heterogeneous terminals,an efficient and aggregated network communication topology structure is adaptively constructed to improve distributed training efficiency and communication bandwidth utilization.(3)A distributed training system for mobile terminals is designed and implemented.Based on the above training framework and network topology generation strategy,a distributed training system for efficient aggregated communication under mobile edge computing is implemented,and the feasibility of the system is verified through functional testing and performance analysis. |
