文献类型：期刊文献

中文题名：结合元学习的去中心化联邦增量学习方法

英文题名：Decentralized Federated Continual Learning Method Combined with Meta-learning

作者：黄楠[1];李冬冬[1];姚佳[1];王喆[1]

机构：[1]华东理工大学信息科学与工程学院,上海200237

年份：2024

卷号：51

期号：3

起止页码：271

中文期刊名：计算机科学

外文期刊名：Computer Science

收录：CSTPCD;;北大核心:【北大核心2023】；CSCD:【CSCD_E2023_2024】；

基金：上海市科技计划(21511100800);国家自然科学基金(62276098)。

语种：中文

中文关键词：去中心化联邦学习;数据安全;增量学习;元学习

外文关键词：Decentralized federated learning;Datasecurity;Continual learning;Meta learning

摘要：针对联邦增量场景中持续学习和数据安全的问题,构建了结合元学习的去中心化联邦增量学习框架。首先,为解决增量场景中持续学习带来的灾难性遗忘问题,提出了结合最近类均值样本回放的增量元学习方法NMR-cMAML,利用元训练对不同任务流的快速适应进行元更新,得到适用于新旧样本的模型。然后,为解决联邦增量场景中的数据安全问题,设计了基于对等网络架构的去中心化联邦增量学习框架,对等架构中每个客户端采用NMR-cMAML对私有的持续任务流进行增量学习。不同于传统的基于服务器-客户端的中心化架构,该去中心化架构采用客户端间通信的策略,消除了传统中央服务器易被攻击的隐患;同时,在联邦通信过程中,通过共享元学习的模型参数实现客户端间知识的有效迁移。最后在图像数据集(Cifar100和Imagenet50)上进行了不同任务场景的实验,结果表明所提方法能在提高系统的数据安全性的同时提升客户端本地性能。
For the problems of continual learning and data security in federated continual scenarios,a decentralized federated continual learning framework combined with meta-learning is constructed.First,in order to solve the problem of catastrophic forgetting in incremental scenarios,an incremental meta-learning method based nearest mean-of-exemplars replaying called NMR-cMAML is proposed.Then,to solve the problem of privacy security in federated continual scenarios,a decentralized federated continual framework based on peer-to-peer network architecture is designed,which is different from the center architecture based on server-client.Each client in the decentralized framework adopts NMR-cMAML to learn the continuous tasks incrementally,and the effective knowledge migration between clients is realized by sharing the meta-learning model in the federal communication process.Finally,experiments are conducted on image data sets(Cifar100 and Imagenet50)to verify that the proposed method improves the data privacy security of the system and improves the local performance of the client.