Federated Noisy Client Learning.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Privacy

Background:

• Federated learning (FL) enables collaborative model training while preserving data privacy through decentralized data.
• Standard FL methods are vulnerable to performance degradation caused by noisy clients and their data.

Purpose of the Study:

• To investigate the impact of noisy clients on federated learning model convergence and performance.
• To propose a novel framework, Federated Noisy Client Learning (Fed-NCL), for robust federated learning with noisy clients.

Main Methods:

• Quantified the negative impact of noisy clients on learned representations across different model layers.
• Developed Fed-NCL to identify noisy clients by estimating data quality and model divergence.
• Implemented robust layerwise aggregation and label correction to address data heterogeneity and improve generalization.

Main Results:

• Confirmed that noisy clients significantly impair global model convergence and performance in FL.
• Observed that noisy clients introduce greater bias in deeper layers compared to earlier layers.
• Demonstrated that Fed-NCL enhances the performance of state-of-the-art FL systems in the presence of noisy clients.

Conclusions:

• Noisy clients pose a critical challenge to federated learning, particularly affecting deeper model layers.
• Fed-NCL effectively identifies and mitigates the negative effects of noisy clients, leading to improved model robustness.
• The proposed framework offers a practical solution for building more reliable federated learning systems.