Dynamic parameter screening and hierarchical optimization methods for heterogeneous federated learning

Le résumé fourni par la source

Federated learning is a machine learning solution for distributed entities that aims to build high-quality machine learning models using local data on dispersed end devices while protecting client privacy. However, in federated learning, differences in client computing power and the non-independent and undistributed nature of the data lead to model performance degradation problems. Therefore, this paper proposes a new federated learning algorithm, FIARSE-HL, whose core is to incorporate the dynamic parameter filtering mechanism of the FIARSE algorithm into the client-side hierarchical architecture. The client dynamically filters important parameters to form personalized layers and shared layers, and then uploads only the filtered parameters of the shared layers to the server aggregation. Experimental results show that, compared to existing federated learning algorithms, the algorithm proposed in this paper exhibits good robustness when aggregating datasets of varying degrees of heterogeneity, achieves better training accuracy in different computing power scenarios using the MNIST dataset, and improves model aggregation accuracy.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.