关键词: 关键词: 联邦学习, 边缘计算, 异构性, 客户端漂移, 算力多样性

Abstract: In view of the challenges of non independent and identically distributed data and heterogeneous computing power faced by federated learning in edge computing applications, the concept of local drift variable is introduced to avoid the large deviation in the update direction of the client model caused by non independent and identically distributed data, which will lead to unstable convergence of the model; By adjusting the local model parameters, the local training process is separated from the global aggregation process, and the performance of Federated learning in the training process of non independent and identically distributed data is optimized. In addition, in view of the diversity of edge servers' computing power, a new strategy is proposed: a simplified neural network sub model is divided from the global model and sent to the edge servers with limited computing power for training, while the edge servers with high computing power use the whole global model for training; The parameters trained by the low computing edge server will be uploaded to the cloud server, and the fitting speed of the whole model will be accelerated by freezing some parameters. Combined with the above two methods, a federated learning optimization algorithm (FedLD) based on local drift and diversity of computing power is proposed, which aims to solve the heterogeneous challenges caused by non independent and identically distributed data and diversity of computing power in the application of Federated learning in edge computing. Experimental results show that the proposed FedLD algorithm has faster convergence speed and higher accuracy than FedAvg、Scaffold and FedProx algorithms, and improves the model accuracy by 0.25%, 1.5% and 5% on MNIST, CIFAR-10 and CIFAR-100 datasets, respectively. Compared with the latest FedProcalgorithm, we can find that the FedLD algorithm has lower communication overhead, and the comparison experiment is carried out in the K-Nearest Neighbor algorithm (KNN), Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) network model. The comparison experiment shows that the prediction accuracy of the three models is improved by about 1% after the combination of FedLD algorithm. Keywords: federated learning; edge computing; heterogeneity; client drift; computing power diversity

Key words: Keywords: federated learning, edge computing, heterogeneity, client drift, computing power diversity