An Efficient Group Federated Learning Framework for Large-Scale EEG-Based Driver Drowsiness Detection.

Area of Science:

• Neuroscience
• Machine Learning
• Transportation Safety

Background:

• Driver drowsiness is a major cause of traffic accidents.
• Monitoring electroencephalogram (EEG) signals offers an effective solution for drowsiness detection.
• Existing methods face challenges with data privacy and insufficient data utilization.

Purpose of the Study:

• To propose a Group Federated Learning (Group-FL) framework for large-scale, privacy-preserving driver drowsiness detection.
• To address data heterogeneity and privacy concerns in driver monitoring systems.
• To improve the efficiency and effectiveness of drowsiness detection using distributed EEG data.

Main Methods:

• A Group Federated Learning (Group-FL) framework was developed, organizing clients into hierarchical groups for efficient aggregation.
• A global-personalized deep neural network was designed to extract shared and fine-grained features from diverse EEG data.
• Three checking modules were implemented to handle data imbalance, pollution, and personalized model application.

Main Results:

• The Group-FL framework demonstrated effective utilization of diverse client data while preserving privacy.
• The global-personalized deep neural network achieved a mean accuracy of 81.0%, F1-score of 82.0%, and AUC of 87.9%.
• Experimental validation confirmed the effectiveness of individual components within the framework.

Conclusions:

• The proposed Group-FL framework offers an efficient and privacy-preserving solution for large-scale driver drowsiness detection.
• The global-personalized deep neural network effectively handles variations in EEG signals across clients.
• This approach significantly advances the potential for real-world application of EEG-based driver monitoring systems.