Cifar10

Federated learning is a distributed machine learning paradigm where a global server iteratively aggregates model parameters of local users without directly accessing their data. However, the heterogeneity of user models and Non-Independently and Identically Distributed (non-IID) data poses significant challenges to federated learning, potentially leading to high communication costs for model parameters and degraded model performance. To address these challenges, this paper proposes a novel heterogeneous federated learning method based on trainable class prototypes, termed FedTCP. The method addresses model heterogeneity and ensures user privacy by sharing only class prototypes among heterogeneous clients. Additionally, a generative model is introduced to enable the trainability of global class prototypes, and contrastive learning is utilized to adaptively enlarge the boundary distance between different class prototypes. Guided by the global class prototypes, users conduct model training by integrating contrastive learning with class prototype learning, thereby enhancing the separability of class prototypes, preserving semantic information, and effectively mitigating performance degradation caused by data heterogeneity. Extensive experiments on various heterogeneous federated frameworks demonstrate that FedTCP significantly outperforms state-of-the-art methods in terms of accuracy while maintaining advantages in communication efficiency and privacy preservation.