FedNTProto: A Prototype-Based Approach for Personalized Federated Learning

Federated learning (FL) is critically important in the field of machine learning and data privacy due to its decentral-ized learning approach, which allows for model training across distributed datasets without sharing raw data. Despite its benefits, FL faces several challenges, including high communication costs, system heterogeneity, statistical heterogeneity, and privacy concerns. This work addresses the issues of communication cost and device heterogeneity by introducing a Federated Prototype Learning approach, which leverages deep metric learning to optimize the distance between global prototypes and local prototypes, effectively reducing communication requirements. Experimental results demonstrate a reduction in communication costs by 100 to 500 times while maintaining acceptable accuracy levels. Moreover, this approach supports the training of different models concurrently within the same FL system, showcasing its robustness and adaptability in heterogeneous environments.