A federated learning method based on class prototype guided classifier for long-tailed data

In federated learning, training on long-tailed data frequently leads to biased classifiers due to a significant imbalance in the number of samples between majority and minority classes. Prototype-based methods have been proven effective in capturing underlying representations in federated learning, contributing to performance improvements by recent studies. However, the class prototypes can be influenced by sample size, class size, and the number of iterations. In this work, we propose a prototype-based federated learning method for long-tailed data by retraining classifiers. This involves freezing the global prototypes aggregated from local prototypes and using them as a regularization term to guide local training. Compared to previous prototype-based methods, our approach focuses on the expression differences of class prototypes in long-tailed data, reduces the classifier's reliance on the majority class samples, and can concentrate more on minority classes. Notably, our method does not introduce extra parameters and communication costs. We conduct experiments on image classification tasks under various settings, and our method outperforms all baselines in terms of performance.