SPARK: A High-Efficiency Black-Box Domain Adaptation Framework for Source Privacy-Preserving Drowsiness Detection

Area of Science:

• Neuroscience
• Computer Science
• Road Safety

Background:

• Electroencephalography (EEG)-based drowsiness monitoring is vital for road safety.
• Existing domain adaptation methods for EEG analysis often require source data or models, raising privacy concerns.
• Cross-subject evaluation is essential for generalizable EEG drowsiness detection.

Purpose of the Study:

• To introduce a novel black-box domain adaptation (BBDA) framework for EEG-based drowsiness detection.
• To address privacy risks associated with traditional unsupervised and source-free domain adaptation methods.
• To develop an efficient and robust system for real-world drowsiness monitoring.

Main Methods:

• Proposed the Self-distillation and Pseudo-labelling for Ensemble Deep Random Vector Functional Link (edRVFL)-based Black-box Knowledge Adaptation (SPARK) framework.
• Utilized entropy-based selection for high-confidence sample pseudo-labeling.
• Employed ensemble self-distillation for knowledge extraction and student model refinement.

Main Results:

• SPARK demonstrated superior performance compared to strong baselines on two public driver drowsiness datasets.
• The framework significantly reduced training time, enhancing computational efficiency.
• The edRVFL features contributed to efficiency, particularly for small datasets.

Conclusions:

• The SPARK framework offers an effective and privacy-preserving solution for EEG-based drowsiness detection.
• SPARK shows potential for practical integration into real-world monitoring systems.
• The study highlights the feasibility of BBDA for sensitive physiological data analysis.