Collaborative Network and Metric Learning Based Label Noise Robust Federated LearningMethod

Currently,there is limited research on the problem of label noise in federated learning.The main approaches involve introducing a benchmark dataset on the server side to evaluate the client's model,controlling the aggregation weights and feature class centers of the clients.However,most methods still have room for improvement in distinguishing noisy clients or noisy samples.This paper proposes a label-noise robust federated learning method based on co-networks and metric learning.The method consists of the following three parts:1) Client mutual evaluation mechanism.Clients score each other's models,construct a rating matrix,and further transform it into an adjacency matrix to differentiate clean/noisy clients.2) Collaborative network module.By constructing two collaborative equivalent federated network models,the Jensen-Shannon divergence is used to distinguish clean samples from noisy samples for the training of collaborative networks.3) Federated-collaborative network triplet loss.A loss function is designed to constrain the output features of the collaborative networks for the same noisy samples.Experimental verification is conducted on the publicly available datasets CIFAR-10 and CIFAR-100,and the results demonstrate the superiority of the proposed method in accuracy.