基于本地模型质量的客户端选择方法

Client Selection Method Based on Local Model Quality

联邦学习是一种针对数据分布于多个客户端的环境下,客户端共同协作训练模型的分布式机器学习方法。在理想情况下全部客户端均参与每轮训练,但是实际应用中只随机选择一部分客户端参与。随机选择的客户端通常不能全面反映全局数据分布特征,导致全局模型训练效率和模型精度降低。为此,提出一种基于本地模型质量的客户端选择方法ChFL。分析影响模型精度和收敛速度的重要因素,提取可反映客户端模型质量的损失值和训练时间2个重要指标。通过对本地损失值和训练时间融合建模,用于评估客户端模型质量。在此基础上,基于客户端质量指导客户端选择,同时与随机选择策略进行一定比例的结合,以提高全局模型精度。通过选择具有高质量的数据且计算性能较优的客户端参与训练,提升模型精度并加快收敛速度。在FEMNIST、CIFAR-10、MNIST、CINIC-10和EMNIST数据集上的实验结果表明,相比3种基线算法FedAvg、FedProx、FedNova,将ChFL与基线算法相结合后的收敛速度平均加快约10%,准确率平均提高4个百分点。

Federated learning is a distributed machine learning method that targets environments where data are distributed across multiple clients that collaborate to train models.In an ideal scenario,all clients participate in each round of training,but in practical applications,only a random portion of clients are selected to participate.Randomly selected clients often fail to fully reflect the global data distribution characteristics,resulting in a decrease in global model training efficiency and model accuracy.ChFL,a client selection method based on local model quality,is proposed.The important factors that affect the accuracy and convergence speed of the model are analyzed,and loss value and training time,two important indicators that can reflect the quality of the client model,are extracted.The combination of local loss values and training time for modeling are used to evaluate the quality of client models.Consequently,client selection is guided based on client quality and combined with a random selection strategy in a certain proportion to improve the accuracy of the global model.By selecting clients with high-quality data and better computational performance to participate in training,model accuracy is improved and convergence speed is accelerated.The experimental results on the FEMNIST,CIFAR-10,MNIST,CINIC-10,and EMNIST datasets show that compared to the three baseline algorithms(FedAvg,FedProx,and FedNova),combining ChFL with the baseline algorithm results in an average convergence speed acceleration of about 10%and an accuracy improvement of about 4 percentage points.