Robust two stages federated learning for sensor based human activity recognition with label noise.

Area of Science:

• Machine Learning
• Artificial Intelligence
• Data Science

Background:

• Federated learning (FL) enables collaborative model training across devices.
• Human activity recognition (HAR) models in FL are hindered by label noise from data annotation.
• Existing methods for label noise in FL have limitations in client evaluation and data aggregation.

Purpose of the Study:

• To propose LN-FHAR, a novel two-stage federated learning framework.
• To enhance label noise robustness and mitigate feature drift in HAR models.
• To address data heterogeneity and improve model generalization in noisy environments.

Main Methods:

• Client selection using class-level loss analysis and Gaussian Mixture Model.
• Noise-robust training with reliable neighbor sample filtering and prototype regularization.
• Data-aware aggregation weighting client contributions by data quality and quantity.

Main Results:

• LN-FHAR effectively mitigates the coupling of label noise and data heterogeneity.
• The framework demonstrates robustness and generalization in complex noisy environments.
• Improved performance in federated human activity recognition models under label noise.

Conclusions:

• LN-FHAR offers a robust solution for federated human activity recognition with label noise.
• The proposed methods enhance client quality assessment and data aggregation strategies.
• This framework advances the development of reliable AI systems in real-world, noisy data scenarios.