Data-Quality-based Aggregation Methods in Federated Learning: a Comprehensive Study

Federated Learning (FL) enables collaborative model training across decentralized devices while preserving data privacy. However, the performance of FL systems is significantly influenced by the varying quality of data residing on individual devices. Data heterogeneity can hinder model convergence and accuracy. Traditional aggregation methods often treat all client updates equally, making them susceptible to the influence of low-quality data. This paper investigates the critical role of data quality in FL and explores the effectiveness of various data quality-based aggregation methods. We examine techniques that weigh client updates based on estimated data quality metrics. We also analyse adaptive aggregation strategies that dynamically adjust the aggregation process based on observed data quality characteristics. Through a comprehensive case study involving image classification tasks on simulated and real-world datasets, we evaluate and discuss the performance of these methods under diverse scenarios. This research provides valuable insights into the challenges posed by data quality and offers a comprehensive overview for developing more robust data quality-based aggregation methods in FL.