Learning to be Fair: A Federated Learning Aggregation Approach on Non-IID Data
Ziyuan Lin, Ling Li · 2023
Federated Learning (FL) enables machine learning on distributed data without sharing raw data. However, data heterogeneity among clients (Non-IID) can lead to unfair and unstable global models. In this paper, we propose FedCA, a novel Federated aggregation algorithm based on Contribution Assessment. FedCA leverages gradient similarity and data quality to assign appropriate weights to clients during model aggregation, aiming to achieve fairness on Non-IID data heterogeneity in the federated learning process. FedCA outperforms existing methods, promoting fair and robust global models in distributed machine learning.