个性化联邦学习的相关方法与展望  

Methods and Prospects of Personalized Federated Learning

作　　者：孙艳华 王子航[1,2] 刘畅 杨睿哲 李萌[1,2] 王朱伟 SUN Yanhua;WANG Zihang;LIU Chang;YANG Ruizhe;LI Meng;WANG Zhuwei(Faculty of Information Technology,Beijing University of Technology,Beijing 100124,China;Advanced Information Network Beijing Laboratory,Beijing University of Technology,Beijing 100124,China)

机构地区：[1]北京工业大学信息学部,北京100124 [2]北京工业大学先进信息网络北京实验室,北京100124

出　　处：《计算机工程与应用》2024年第20期68-83,共16页Computer Engineering and Applications

基　　金：北京市自然科学基金(4222002)。

摘　　要：目前,随着人工智能研究的进步,人工智能被大规模采用,数据监管等领域的需求也促使人们对隐私保护的认识和关注越来越多,这促进了联邦学习(federated learning,FL)框架的流行。但现有的FL难以应对异构问题以及用户的个性化需求。为了应对上述问题,研究了个性化联邦学习(personalized federated learning,PFL)的相关方法并提出了展望。列举了FL的框架并指出了FL的不足,在FL场景的基础上,引出PFL的研究动机对PFL中的统计异构、模型异构、通信异构、设备异构进行分析并提出可行性方案;将PFL中的客户端选择、知识蒸馏等个性化算法分类并分析各自的创新与不足。最后,对PFL的未来研究方向进行了展望。Currently,with the advancement of artificial intelligence research,artificial intelligence is being widely adopted,and the increasing demand in areas such as data governance has led to growing awareness and concern for privacy protection,this has promoted the popularity of the federated learning(FL)framework.However,existing FL frameworks struggle to address heterogeneous issues and personalized user needs.In response to these challenges,methods of personalized federated learning(PFL)are studied and prospects are proposed.Firstly,the FL framework is outlined and its limitations are identified,leading to the research motivation for PFL based on FL scenarios.Subsequently,the analysis of statistical heterogeneity,model heterogeneity,communication heterogeneity,and device heterogeneity in PFL is conducted,and feasible solutions are proposed.Then,personalized algorithms in PFL such as client selection and knowledge distillation are categorized,and their innovations and shortcomings are analyzed.Finally,future research directions for PFL are discussed.

关 键 词：个性化联邦学习(PFL) 数据监管 异构问题 隐私保护 

分 类 号：TP181[自动化与计算机技术—控制理论与控制工程]