基于模糊推理和熵约束的联邦半监督学习算法OA

This study tackled model bias in federal semi-supervised learning(FSSL)caused by scarce labeled data and hetero-geneous data distributions.The research proposed a new learning scheme that used fuzzy inference and entropy constraints.Most existing methods generate pseudo-labels from predictions to leverage unlabeled data.They usually set confidence thres-holds to screen these predictions.However,these approaches often missed the deeper reasons behind low confidence scores.Unlabeled clients can show significant differences in their data distributions.Using noisy pseudo-labels without distinction de-grades model performance.To resolve this problem,this paper designed a fuzzy computation module to model uncertainty.This module explored potential information in low-confidence pseudo-labels,paying special attention to pseudo-labels from con-fusing categories.The scheme also introduced a filtering mechanism to use information entropy.This mechanism effectively removed low-quality unlabeled data and reduced the impact of noisy samples.The research created FedFREC,a cascaded learning algorithm for FSSL.The algorithm combined the fuzzy computation module for confusing categories with the entropy-based filtering mechanism for low-confidence samples.This approach successfully utilized the information embedded in unla-beled data.Extensive experiments on the SVHN,CIFAR-10,CIFAR-100,fashion MNIST,and ISIC2018 datasets demon-strate the superiority of the proposed method.