Research Article
FedCL: An Efficient Federated Unsupervised Learning for Model Sharing in IoT
@INPROCEEDINGS{10.1007/978-3-031-24383-7_7, author={Chen Zhao and Zhipeng Gao and Qian Wang and Zijia Mo and Xinlei Yu}, title={FedCL: An Efficient Federated Unsupervised Learning for Model Sharing in IoT}, proceedings={Collaborative Computing: Networking, Applications and Worksharing. 18th EAI International Conference, CollaborateCom 2022, Hangzhou, China, October 15-16, 2022, Proceedings, Part I}, proceedings_a={COLLABORATECOM}, year={2023}, month={1}, keywords={Federated learning Internet of things Self-supervised learning Unsupervised learning}, doi={10.1007/978-3-031-24383-7_7} }- Chen Zhao
Zhipeng Gao
Qian Wang
Zijia Mo
Xinlei Yu
Year: 2023
FedCL: An Efficient Federated Unsupervised Learning for Model Sharing in IoT
COLLABORATECOM
Springer
DOI: 10.1007/978-3-031-24383-7_7
Abstract
Federated Learning (FL) continues to make significant advances, solving model sharing under privacy-preserving. However, these existing methods are only of limited utility in the Internet of Things (IoT) scenarios, as they either heavily depend on high-quality labeled data or only perform well under idealized conditions, which typically cannot be found in practical applications. As such, a natural problem is how to leverage unlabeled data among multiple clients to optimize sharing model. To address this shortcoming, we propose Federated Contrastive Learning (FedCL), an efficient federated learning method for unsupervised image classification. The proposed FedCL can be summarized in three steps: distributed federated pretraining of the local model using contrastive learning, supervised fine-tuning on a server with few labeled data, and distillation with unlabeled examples on each client for refining and transferring the personalized-specific knowledge. Extensive experiments show that our method outperforms all baseline methods by large margins, including 69.32% top-1 accuracy on CIFAR-10, 85.75% on SVHN, and 74.64% on Mini-ImageNet with the only use of 1% labels.
