Semi-Supervised Federated Learning over Relevant Heterogeneous Data

Federated Learning (FL) aims to collaboratively train models without clients’ data transfer. However, FL faces significant challenges, particularly concerning model performance, which heavily relies on the quality of clients’ data. Clients’ data quality varies due to factors like data heterogeneity and labeling strategies across clients. Not all clients have ground truth labels; client data are often unlabeled, partially labeled, or labeled with sub-optimal quality due to labeling cost (time and resources) or insufficient expertise. This paper introduces a novel approach based on Semi-Supervised FL (SSFL), which employs a Multi-Purpose pre-trained SSFL model (MP-SSFL) to leverage ground truth labels from server data for training a global pseudo-labeling model. Such model is, in turn, utilized for label prediction and distributed data selection across clients while ensuring privacy preservation. Our MP-SSFL method enables labeling of unlabeled data, correction of inaccurately labeled data, and mitigation of label overlapping issues. Compared against benchmarks, MP-SSFL significantly enhances the quality of pseudo-labels, thereby leading to improved performance of global classifier models. Such improvement is notable in cases involving non-identically labeled data across clients with non-independently and identically distributed data.