关键词: 联邦学习, 非独立同分布, 余弦相似度, 正则化, 个性化联邦学习, 隐私安全

Abstract: Federated learning scenarios often face the problem of data heterogeneity and the need to provide personalized models for different task requirements. However, there is a trade-off between personalization and global generalization in some existing personalized federated learning. Most of them use traditional FL based on client data. Quantity-weighted aggregation, aggregation weights lack personalization. In response to the above problems, a personalized federated learning method (pFedSCR) based on similarity clustering and regularization is proposed. The pFedSCR algorithm trains the private model in the client's local update phase, introduces L2 norm regularization, and dynamically controls the extent to which the private model refers to the global model; in the server aggregation phase, it clusters based on the client model similarity and dynamically adjusts the client aggregation weight as Different clients aggregate personalized models. Experimental results show that non-IID data scenarios are simulated through Dirichlet distribution on three data sets such as CIFAR-10. Compared with five algorithms including the classic algorithm Fedprox and the latest personalized algorithm fedpcl, the pFedSCR algorithm can perform well in various scenarios. Get higher accuracy, up to 99.03% test accuracy.

Key words: Federated Learning (FL), Non-Independent Identical Distribution (Non-IID), Cosine Similarity, regularization, Personalized Federated Learning(PFL), privacy security