Currently,federated semi-supervised learning(FSSL)faces the challenge of making effective use of a large amount of unlabeled data during training.Although knowledge sharing between clients through a lightweight prototyp-ing network can alleviate pseudo-label quality issues,there are still bottlenecks.In this paper,we propose a federated semi-supervised learning model based on dynamic threshold enhanced prototype network.By introducing Curriculum Pseudo labeling,the core is to dynamically adjust the threshold of the learning state of different classes of samples,so that the model can learn high-quality samples and significantly improve the prediction performance of the model.Exper-imental results show that our proposal has achieved excellent test performance on multiple datasets.On the CIFAR-10 dataset,our proposal improves the test accuracy by at least 3%compared with similar algorithms.In addition,there is a 1%～7%lead on SVHN and STL-10 datasets.It is worth noting that our proposal performs well in handling heterogen-eous and homogeneous data,and has good adaptability to different proportions of labeled and unlabeled data.Our pro-posal can improve the test accuracy.What's more,it does not add additional communication overhead and computation-al cost.These results suggest that our proposal has great potential in the field of federated semi-supervised learning,and provides a high-performance and high-efficiency solution for practical applications.
关键词
联邦学习/半监督学习/知识共享/原型网络/伪标签/动态阈值/无标签数据/数据异质性
Key words
federated learning/semi-supervised learning/knowledge sharing/prototypical network/pseudo label/dy-namic threshold/unlabeled data/heterogeneous data
维普
万方数据