FCAT⁃FL:基于Non⁃IID数据的高效联邦学习算法

FCAT⁃FL:an efficient federated learning algorithm based on Non⁃IID data

针对非独立同分布(Non⁃IID)数据影响联邦学习收敛速度、公平性和准确性的问题,提出一种基于Non⁃IID数据的快速收敛公平联邦迁移学习框架———FCAT⁃FL。该框架改进传统联邦学习依照客户端数据量占比权衡聚合贡献度的策略,根据客户端本地模型参数和服务器聚合模型参数间的关系,在每轮聚合时为客户端动态分配自适应权重,并在客户端引入个性化迁移学习模型和动量梯度下降算法以求加快本地模型训练速度。实验结果表明:与几种基线聚合策略相比,当部分客户端的数据为Non⁃IID时,FCAT⁃FL中聚合策略1的全局迭代轮次有所减少,客户端间公平性和准确性得到提高,并且迁移学习的使用令客户端需训练和上传的模型参数数量减少,使FCAT⁃FL适用于客户端资源有限的移动边缘网络。

Aiming at the problem that non⁃independent and identically distributed(Non⁃IID)data affects the convergence speed,fairness and accuracy of federated learning,a fast and fair federated migration learning framework—FCAT⁃FL based on Non⁃IID data is proposed.This framework improves the traditional federated learning strategy of weighing the contribution of aggregation according to the proportion of client data,and assigns adaptive weights to clients in each round of aggregation dynamically according to the relationship between client model parameters and server model parameters.And a personalized migration learning model and a momentum gradient descent algorithm are introduced in the client to speed up the convergence of the local model.The experimental results show that,compared with several baseline aggregation strategies,when the data of some clients is Non⁃IID,the global iteration round of strategy 1 in FCAT⁃FL is reduced and the fairness and the accuracy among clients are improved.Moreover,the use of migration learning reduces the number of model parameters that clients need to train and upload.Therefore,FCAT⁃FL is suitable for mobile edge networks with limited client resources.