Cycle-Fed: A Double-Confidence Unlabeled Data Augmentation Method Based on Semisupervised Federated Learning

The Internet of Things (IoT) generates a substantial volume of unlabeled personal privacy data in finance and healthcare, distributed across diverse locations and networks, which is currently underutilized. Semisupervised federated learning emerges as a promising solution by conducting model training on local devices without transmitting raw data to a central server. This approach enhances the security and efficiency of IoT systems. First, to overcome traditional data augmentation limitations in simulating raw data distribution, we introduce a data augmentation module using a uniformly distributed dropout (Uout) layer. This module enhances data diversity by mitigating sensitivity to variance shifts. Furthermore, considering the insufficiency of pseudo-label availability under the condition of low-density separation, we propose the Cycle-Fed model with dual-reliability. This model enhances its performance through the effects of data augmentation by incorporating pseudo-labeled positive samples subjected to secondary validation by discriminators provided by the data augmentation module. Finally, we propose a client-side optimal value avoidance strategy based on an adaptive local proximal term, which is denoted as $\mu _{t}$ . Experimental results on a public dataset indicate that the Cycle-Fed model surpasses the baseline with a 4.52%-8.76% reduction in loss, 1.62%-6.09% accuracy improvement, and a 1.285%-1.396% increase in area under the curve.