Anti-noise robust heterogeneous federated learning method

The invention discloses an anti-noise robust heterogeneous federated learning method. According to the method, firstly, the feedback of the models is directly adjusted by utilizing common data, communication between heterogeneous models can be completed without cooperation of a global model, secondly, the negative influence of label noise in a client is reduced by applying a robust noise-resistant loss function, and finally, the noise feedback from other clients is effectively reduced. The invention provides a client confidence re-weighting scheme. According to the scheme, a corresponding weight is adaptively allocated to each client in a collaborative learning stage. According to the method, the negative influence of various types and proportions of noise in a noise heterogeneous client environment can be effectively reduced, so that the performance of the model is effectively improved. 本发明公开了一种抗噪声的鲁棒异构联邦学习方法。首先，本发明通过利用公共数据直接调整模型的反馈，不需要全局模型的协作就可以完成异构模型之间的通信，其次，本发明应用一个稳健的耐噪声损失函数来减少客户端内部标签噪声的负面影响，最后，对于来自其他客户端的