关键词: 关键词: 个性化联邦学习, 数据异构, 双流网络, 新客户端问题

Abstract: Abstract: Classic federated learning algorithms are difficult to achieve good results in scenarios where data is highly heterogeneous and unevenly distributed. Personalized federated learning proposes a new solution to the problem of data heterogeneity in federated learning, which is to "tailor" a dedicated model for each client, which has good performance. However, this raises the issue of difficulty in extending federated learning to new clients. This paper focuses on the challenges of performance and scalability in personalized federated learning, and proposes a federated learning model based on a dual stream network structure. The dual stream network model can not only have the performance of personalized models but also be easily extended to new clients by adding an encoder for analyzing the clients. Experiments have shown that in dataset tasks such as MNIST, CIFAR10, and FashionMNIST, there is an improvement of more than ten percentage points compared to the classic federated average algorithm, while also solving the problem of difficult scalability for new clients.

Key words: personalized federated learning, data heterogeneous, dual stream neural network , new client