TSFAN: tensorized spatial-frequency attention network with domain adaptation for cross-session EEG-based biometric...

Area of Science:

• Biometrics
• Neuroscience
• Machine Learning

Background:

• Electroencephalogram (EEG) signals offer promising, invisible biometric features for high-security applications.
• Challenges exist in EEG biometrics due to device and subject state variations across sessions, impacting data distribution.
• Existing multi-source unsupervised domain adaptation (MUDA) methods often neglect relationships among domain-invariant features.

Purpose of the Study:

• To propose a novel MUDA method, Tensorized Spatial-Frequency Attention Network (TSFAN), to enhance EEG-based biometric recognition performance in target domains.
• To model significant relationships of domain-invariant features using a tensorized attention mechanism.
• To address the curse of dimensionality and ensure scalability with an arbitrary number of sessions.

Main Methods:

• Developed TSFAN, a MUDA method utilizing a tensorized attention mechanism to model common spatial-frequency representations across source and target domains.
• Employed Tucker format approximation for TSFAN to manage dimensionality and achieve linear scalability with the number of domains.
• Validated TSFAN's effectiveness through extensive experiments on representative benchmarks using cross-session validation.

Main Results:

• TSFAN demonstrated superior performance in EEG-based biometric recognition compared to state-of-the-art approaches.
• Cross-session validation confirmed the effectiveness of TSFAN in handling domain shifts.
• Electrode selection analysis indicated that stable EEG identity features are distributed across brain regions, with 20 electrodes from the 10-20 system being sufficient.

Conclusions:

• TSFAN successfully investigates consistent EEG identity features across sessions by leveraging a novel tensorized attention mechanism.
• The method effectively collaborates intra-source transferable information with inter-source interactions, remaining robust to domain shifts.
• EEG biometrics can achieve stable identity recognition across sessions by utilizing specific brain region electrode data.