基于动态阈值增强原型网络的联邦半监督学习模型

Federated semi-supervised learning model based on dynamic threshold enhanced prototype network

目前,联邦半监督学习面临着有效利用训练过程中大量无标签数据的挑战。尽管通过轻量级的原型网络实现客户端之间的知识共享可以缓解伪标签质量问题,但仍然有瓶颈。本文提出一种新的动态阈值增强下的原型网络联邦半监督学习算法。通过引入课程伪标签技术,其核心是对不同类别样本的学习状态动态调整阈值,使模型能够学习高质量的样本,显著提高模型的预测性能。实验结果表明,本算法在多个数据集上均取得优异的测试性能。在CIFAR-10数据集上,本算法相对于同类算法至少提高3%的测试精度。此外在SVHN和STL-10数据集上也有1%~7%的领先优势。值得注意的是,本算法在处理异质性和同质性数据时表现出色,且对于不同比例的有标签和无标签数据都具有良好的适应性。本算法不仅提高测试精度,而且未带来额外的通信开销和计算成本。这些结果表明本算法在联邦半监督学习领域具有巨大潜力,并为实际应用提供了一个性能卓越且高效的解决方案。

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 prototyping 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.Experimental 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 heterogeneous and homogeneous data,and has good adaptability to different proportions of labeled and unlabeled data.Our proposal can improve the test accuracy.What’s more,it does not add additional communication overhead and computational 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.