Federated semi-supervised learning with contrastive representations against noisy labels

Federated semi-supervised learning presents a pragmatic scenario wherein a centralized model is trained utilizing a server with access to labeled data, while participating clients lack any labeled data. In this context, the inaccuracy of real-world labels on the server available for training poses a huge challenge to the federated semi-supervised learning. These inaccuracies can have a detrimental impact on the overall performance of the system and impose limitations on its use. In this paper, we propose a novel Federated Semi-supervised learning framework with Contrastive Representations, called FedCR, with the aim of addressing the aforementioned ubiquitous problems in the field of image classification tasks. Firstly, our approach employs contrastive representation learning to build memory representations of images, which can learn an image’s general features from an augmented view without relying on negative pairs and prevent the model from memorizing noise. Then we take a cautious approach during model updates to prevent any potential leakage to ensure the privacy and security of the clients’ information. Additionally, for the sake of improving robustness of the model, a contrastive regularization function is applied to preserve information connected to true labels while filtering out information associated with wrong labels. Furthermore, we mitigate the negative impact of mislabeled data during supervised learning by utilizing an improved cross-entropy loss function. Extensive experiments on prevalent datasets for image classification tasks show that the proposed method surpasses previously established state-of-the-art federated semi-supervised learning algorithms and efficiently alleviates the issue of model over-fitting to erroneous labels, especially when label noise is present.