Federated Long-Tailed Learning by Retraining the Biased Classifier with Prototypes

Federated learning is a privacy-preserving framework that collaboratively trains the global model without sharing raw data among clients. However, one significant issue encountered in federated learning is that biased classifiers affect the classification performance of the global model, especially when training on long-tailed data. Retraining the classifier on balanced datasets requires sharing the client's information and poses the risk of privacy leakage. We propose a method for retraining the biased classifier using prototypes, that leverage the comparison of distances between local and global prototypes to guide the local training process. We conduct experiments on CIFAR-10-LT and CIFAR-100-LT, and our approach outperforms the accuracy of baseline methods, with accuracy improvements of up to 10%.