Federated Data Quality Assessment Approach: Robust Learning With Mixed Label Noise

Area of Science:

• Machine Learning
• Distributed Systems
• Data Privacy

Background:

• Federated learning (FL) enables distributed model training without sharing local data.
• Label noise presents a significant challenge in FL due to inaccessibility of local data.
• Existing methods often fail to address mixed noise types common in real-world scenarios.

Purpose of the Study:

• To propose a novel FL method, FedMIN, for robust training in the presence of mixed label noise.
• To develop a framework capable of discriminating noise types and enhancing model performance.
• To improve the accuracy and reliability of federated models under noisy data conditions.

Main Methods:

• FedMIN utilizes a composite framework to model generalized noise patterns by capturing local-global distribution differences.
• Adaptive thresholds are determined for identifying mixed label noise on each client.
• Appropriate weights are assigned during model aggregation, and a loss alignment mechanism using Gaussian Mixture Models (GMMs) is incorporated.

Main Results:

• FedMIN demonstrates superior noise estimation capabilities, leading to improved global model performance.
• Experiments on simulated (CIFAR-10, CIFAR-100, SVHN) and real-world (Camelyon17, MoNuSAC) datasets show significant accuracy gains.
• FedMIN achieved up to a 9.9% improvement in model accuracy compared to existing FL benchmarks.

Conclusions:

• FedMIN effectively addresses the challenge of mixed label noise in federated learning.
• The proposed method enhances the robustness and accuracy of federated models in distributed, noisy environments.
• FedMIN offers a promising solution for real-world applications where data quality is variable.