MINDFL: Mitigating the Impact of Imbalanced and Noisy-labeled Data in Federated Learning with Quality and Fairness-Aware Client Selection
Chaoyu Zhang, Ning Wang, Shanghao Shi, Changlai Du, Wenjing Lou, Y. Thomas Hou · 2023
Federated Learning (FL) has been acclaimed for enhancing machine learning privacy, but it also faces criticism for its high communication overhead due to the need for a large number of interactions between participants and the server. The conventional approach of randomly selecting clients for FL participation can reduce communication costs, but it often slows down the convergence of the global model and lowering its overall quality. Although various advanced methods have been proposed for client selection, focusing on the selection of the most informative local models, these solutions frequently assume clean training data. They often overlook the impact of noisy-labeled and imbalanced local data among clients, which can significantly hinder training efficiency.To address these challenges, we present MINDFL, a client selection mechanism that prioritizes quality and fairness considerations. MINDFL tackles the problem of utility divergence among local models caused by noisy-labeled and imbalanced data. It also ensures fairness in the selection process. In particular, we introduce a novel metric called Quality-of-Model (QoM), which assesses the contribution of local models to the aggregated global model. MINDFL selects clients with high QoM, representing the most informative model updates, in each FL round to maximize learning efficiency. Further, to enhance participant diversity while maintaining clients’ model quality, we utilize a sortition-inspired selection method to choose clients with high QoM randomly. Our comprehensive experimental evaluations demonstrate the effectiveness of MINDFL in improving learning speed and reducing communication overhead.