车联网联邦学习的数据异质性问题及基于个性化的解决方法综述

Survey on data heterogeneity problems and personalization based solutions of federated learning in Internet of vehicles

在车联网(IoV)场景中,不同设备存在海量非独立同分布的数据,容易引发数据异质性问题,影响模型训练性能并威胁交通安全,对此聚焦于车联网联邦学习(FL)的数据异质性问题,通过对问题归因溯源提出了基于个性化的解决方法体系与研究新思路。首先,论述了联邦学习用于车联网的必要性,调研总结了车联网联邦学习中典型的数据异质性问题;其次,从感知、计算和传输3个环节对车联网联邦学习的数据异质性问题进行了分类和追踪;再次,引入个性化方法作为解决各类车联网联邦学习数据异质性问题的核心手段,并分析了现有个性化联邦学习的优点与不足;最后,讨论了个性化联邦学习在车联网场景中面临的研究挑战,并结合无线通信等相关技术展望了未来研究方向。

In Internet of vehicles(IoV)scenario,there was a massive amount of non-independent and identically distributed data among devices,leading to data heterogeneity problems of federated learning(FL).This problem affected the performances of model training and might pose threats to traffic safety.Therefore,the focus lied on the data heterogeneity problem of FL in IoV,the personalized solution system and new research ideas were proposed through problem attribution.Firstly,the necessity of applying FL to IoV was discussed.Through an examination of current applications,identified the data heterogeneity problems of FL in IoV.Secondly,classified and traced the data heterogeneity problems of FL in IoV,from the perspective of perception,computation,and transmission respectively.Thirdly,personalized methods were introduced as the core approaches to address the data heterogeneity problems of FL in IoV,and analyzed the advantages and disadvantages of existing personalized federated learning(PFL).Finally,the challenges encountered by PFL in IoV were outlined,along with the future research prospection related to advanced technologies on wireless communications.