移动边缘计算下基于联邦学习的动态QoS优化  

Dynamic QoS Optimization Method Based on Federal Learning in Mobile Edge Computing

作　　者：张鹏程[1,2] 魏芯淼 金惠颖 ZHANG Peng-Cheng;WEI Xin-Miao;JIN Hui-Ying(School of Computer and Information,Hohai University,Nanjing 211000)

机构地区：[1]水利部水利大数据重点实验室(河海大学),南京211000 [2]河海大学计算机与信息学院,南京211000

出　　处：《计算机学报》2021年第12期2431-2446,共16页Chinese Journal of Computers

基　　金：江苏省自然科学基金(BK20191297);中央高校基本科研业务费专项资金(B210203070,B210202075)资助。

摘　　要：在5G边缘网络飞速发展的过程中,边缘用户对高带宽、低时延的网络服务的质量要求也显著提高.从移动边缘网络的角度来看,网络内的整体服务质量与边缘用户的分配息息相关,用户移动的复杂性为边缘用户分配带来困难,边缘用户分配过程中还存在隐私泄露问题.本文提出一种移动边缘环境下基于联邦学习的动态QoS(Quality of Service)优化方法MECFLD_QoS,基于联邦学习的思想,优化边缘区域的服务缓存,在动态移动场景下根据用户位置分配边缘服务器,有效保护用户隐私,实现区域服务质量优化,对动态用户移动场景有更好的适应性.MECFLD_QoS主要做了以下几个方面的优化工作:(1)优化了传统QoS数据集,将数据集映射到边缘网络环境中,充分考虑边缘计算的移动、分布式、实时性、复杂场景等特点,形成边缘QoS特征数据集;(2)优化了边缘服务器缓存,在用户终端训练用户偏好模型,与区域公有模型交互时只传输参数,将用户的隐私数据封装在用户终端中,避免数据的传输,可以有效地保护用户特征隐私;(3)优化了用户移动场景,在动态移动场景中收集用户移动信息,利用用户接入基站的地理位置拟合用户的移动轨迹进行预测,有效地模糊了用户的真实位置,在轨迹预测的同时有效地保护了用户的位置隐私;(4)优化了用户分配方法,提出改进的基于二维解的人工蜂群算法对边缘网络中的用户分配问题进行优化,事实证明改进的人工蜂群算法针对其多变量多峰值的特点有效地优化了用户分配,达到了较优的分配效果.通过边缘QoS特征数据集实验表明,本方法在多变量多峰值的用户分配问题中能产生全局最优的分配.The development of 5G networks has expanded the Internet of Things,promoted the operation of cellular networks,and further pushed the Internet to the edge of the network.In the process of rapid development of 5G edge networks,mobile edge users have significantly improved the quality of service(QoS)requirements for high-bandwidth and low-latency Web services.Traditional QoS calculation and optimization are from the perspective of service,and the calculation of quality of experience(QoE)is from the perspective of edge users.From the perspective of mobile edge networks,the overall service quality of the edge area is closely related to the edge users allocation(EUA).As a result,EUA needs to be taken into consideration in the optimization work.The allocation of edge users to achieve the best overall service quality in the area is a hot topic.The complexity of user movement behavior and scenes brings difficulties to the edge user allocation.There are also privacy leakage problems in the edge user allocation process.This paper proposes a dynamic QoS optimization method in mobile edge environment based on federal learning(MECFLD_QoS).Based on the essence of federal learning,the proposed method uses gradient descent to solve the logistic regression model,trains the regional public model to optimize the Web service cache of the edge area,and allocates edge servers according to the user location in the dynamic mobile scenario,effectively protecting user privacy.The regional service quality is optimized,and it has better adaptability to dynamic user mobility scenarios.MECFLD_QoS mainly optimized the following aspects:(1)Optimizes the traditional QoS data set,maps the data set to the edge network environment,and fully considers the mobile,distributed,real-time,and complex scenarios features of edge computing,and forms an edge QoS feature data set;(2)Optimizes the edge server cache,trains the user preference model on the user terminals,only transmits model parameters when interacting with the regional public model,and encapsul

关 键 词：移动边缘 联邦学习 移动感知 边缘用户分配 服务质量 

分 类 号：TP18[自动化与计算机技术—控制理论与控制工程]