Noise-Robust Federated Learning via Interclient Co-Distillation.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Distributed Systems

Background:

• Federated learning (FL) enables collaborative model training across multiple clients while preserving user privacy.
• The performance of FL is often hindered by the challenge of obtaining accurately labeled data in real-world scenarios, leading to noisy datasets.
• Robust methods are needed to train shared models effectively using distributed, noisy labeled data.

Purpose of the Study:

• To propose FedDQ, a novel federated learning framework designed for noise-robustness.
• To address the challenge of training high-performance models with distributed noisy labeled data.
• To improve the effectiveness of federated learning in practical applications where data quality is a concern.

Main Methods:

• FedDQ employs a noise-adaptive training strategy to dynamically adjust client training based on estimated label noise levels.
• A co-distillation technique using a two-head network facilitates knowledge transfer and shared representational capabilities among clients.
• An enhanced label correction mechanism, including co-filtering, is utilized to rectify improper labels within the distributed datasets.

Main Results:

• FedDQ demonstrates significant improvements in model performance when handling noisy data in federated learning settings.
• The noise-adaptive strategy effectively mitigates the impact of incorrect labels while leveraging clean data features.
• Experimental results on CIFAR-100 with noisy labels showed up to a 32.4% improvement compared to baseline methods.

Conclusions:

• FedDQ offers an effective solution for robust federated learning with noisy labeled data.
• The proposed co-distillation and quality-aware aggregation techniques enhance model accuracy and reliability.
• This framework advances the practical applicability of federated learning in privacy-sensitive, data-scarce environments.