A deep descriptor for cross-tasking EEG-based recognition.

Area of Science:

• Neuroscience
• Computer Science
• Biometrics

Background:

• Vital signs, including electroencephalogram (EEG) signals, are increasingly used in expert systems and biometrics.
• Motor tasks, whether performed or imagined, significantly alter brain wave patterns and EEG signal characteristics.
• Exploring EEG biometrics from a cross-task perspective is crucial for robust identification systems.

Purpose of the Study:

• To develop a novel method for biometric verification using EEG signals, specifically addressing the challenges posed by different motor tasks.
• To create a deep EEG signal descriptor capable of assessing the impact of motor tasks on biometric accuracy.
• To evaluate the proposed method's performance using a comprehensive dataset and rigorous evaluation protocol.

Main Methods:

• A deep learning approach combining Convolutional Neural Networks (CNNs) and Squeeze-and-Excitation Blocks was developed.
• A data augmentation technique was employed to address the limited data volume for training deep CNN models.
• An evaluation protocol was designed to assess robustness concerning the number of EEG channels and ensure non-overlapping train/test sets.

Main Results:

• The novel method achieved a new state-of-the-art Equal Error Rate (EER) of 0.1% in the cross-task biometric verification scenario.
• Squeeze-and-Excitation-based networks demonstrated superior performance compared to simple CNN architectures in three out of four cross-individual scenarios.
• The proposed data augmentation technique improved the overall performance of the deep CNN model.

Conclusions:

• The developed deep EEG signal descriptor effectively assesses the impact of motor tasks on biometric verification.
• The Squeeze-and-Excitation enhanced CNN approach offers a robust solution for cross-task EEG-based biometrics.
• This research advances the field of EEG biometrics, demonstrating high accuracy and robustness in identifying individuals across varying tasks.