Evidence of Task-Independent Person-Specific Signatures in EEG using Subspace Techniques

Mari Ganesh Kumar, Shrikanth Narayanan, Mriganka Sur, Hema A. Murthy

Abstract

Electroencephalography (EEG) signals are promising as alternatives to other biometrics owing to their protection against spoofing. Previous studies have focused on capturing individual variability by analyzing task/condition-specific EEG. This work attempts to model biometric signatures independent of task/condition by normalizing the associated variance. Toward this goal, the paper extends ideas from subspace-based text-independent speaker recognition and proposes novel modifications for modeling multi-channel EEG data. The proposed techniques assume that biometric information is present in the entire EEG signal and accumulate statistics across time in a high dimensional space. These high dimensional statistics are then projected to a lower dimensional space where the biometric information is preserved. The lower dimensional embeddings obtained using the proposed approach are shown to be task-independent. The best subspace system identifies individuals with accuracies of 86.4% and 35.9% on datasets with 30 and 920 subjects, respectively, using just nine EEG channels. The paper also provides insights into the subspace model's scalability to unseen tasks and individuals during training and the number of channels needed for subspace modeling.

Introduction

While EEG-based person recognition offers high security against spoofing, most existing systems are "task-dependent." A system trained on a subject in a resting state often fails to recognize that same individual during active mental tasks. This limitation arises because EEG signals capture both the brain's state (the task) and its trait (the individual). This research addresses the challenge of extracting a permanent, person-specific "brainprint" that persists across diverse mental activities and recording sessions by treating task-related variance as a nuisance variable to be normalized.

Methods

Drawing inspiration from text-independent speaker recognition—where the goal is to identify a speaker regardless of the words spoken—the authors adapted two state-of-the-art subspace frameworks:

  1. Modified i-vector: A generative approach using Gaussian Mixture Models (GMMs) to estimate high-dimensional statistics (means and variances) across time, projecting them into a lower-dimensional "total variability" space.
  2. Modified x-vector: A discriminative approach using Deep Neural Networks (DNNs). It employs 1D-convolutional layers to extract frame-level features, which are then pooled into a segment-level embedding.
Modified x-vector architecture for multi-channel EEG processing.

The study's primary contribution is the ix-vector, a novel fusion of these two frameworks. To handle multi-channel EEG data effectively, the authors modified the standard architectures to concatenate statistics across channels at an intermediate level, rather than using simple early or late fusion. Finally, Linear Discriminant Analysis (LDA) was applied to the embeddings to maximize the distance between different individuals while minimizing the distance between different tasks for the same person.

Experimental Setup

The robustness of the proposed subspace techniques was evaluated using two distinct datasets:

  • Dataset 1 (Lab-controlled): 30 healthy subjects performing 12 different tasks (e.g., auditory oddball, motor imagery, visual stimuli) across up to 5 sessions using a 128-channel system.
  • Dataset 2 (Clinical/Large-scale): A subset of the Temple University Hospital EEG (TUH-EEG) corpus featuring 920 subjects. This represents one of the largest subject pools in EEG biometric research, providing a rigorous test in noisy, real-world clinical conditions.

Results and Discussion

The combined ix-vector system significantly outperformed baselines, including EEGNet and CNN-RNN models. On Dataset 1, the system achieved 86.4% accuracy and a 5.02% Equal Error Rate (EER) across mismatched tasks. On the 920-subject Dataset 2, it achieved 35.9% accuracy, a remarkably high figure for such a large-scale, noisy classification problem.

The 9-channel EEG configuration used for robust person identification.

Key findings include:

  • Task Independence: The system maintained high accuracy (e.g., ~94% for "Imagining Binary Answers") even when the specific test task was excluded from the training set.
  • Distributed Signatures: Person-specific information is present across the entire scalp. While high-density arrays were used for initial data collection, the researchers found that a strategic subset of 9 channels (covering frontal, central, parietal, and occipital regions) provides an optimal balance between computational efficiency and recognition accuracy.
  • Scalability: The subspace models generalized well; a system trained on one group of individuals could successfully extract unique, meaningful embeddings for entirely new subjects.

Conclusion

This work demonstrates that subspace techniques can effectively normalize task-related variance to reveal stable, person-specific EEG signatures. While performance on clinical data (Dataset 2) highlights the ongoing challenge of artifacts like movement or seizures, the results prove that task-independent biometrics are feasible even at scale. Future research will focus on channel-independent models and the long-term stability of these signatures over several years.

Summary updated

Abstract

Electroencephalography (EEG) signals are promising as alternatives to other biometrics owing to their protection against spoofing. Previous studies have focused on capturing individual variability by analyzing task/condition-specific EEG. This work attempts to model biometric signatures independent of task/condition by normalizing the associated variance. Toward this goal, the paper extends ideas from subspace-based text-independent speaker recognition and proposes novel modifications for modeling multi-channel EEG data. The proposed techniques assume that biometric information is present in the entire EEG signal and accumulate statistics across time in a high dimensional space. These high dimensional statistics are then projected to a lower dimensional space where the biometric information is preserved. The lower dimensional embeddings obtained using the proposed approach are shown to be task-independent. The best subspace system identifies individuals with accuracies of 86.4% and 35.9% on datasets with 30 and 920 subjects, respectively, using just nine EEG channels. The paper also provides insights into the subspace model's scalability to unseen tasks and individuals during training and the number of channels needed for subspace modeling.
BibTeX
@article{kumar2021evidence,
  author = {Mari Ganesh Kumar and Shrikanth Narayanan and Mriganka Sur and Hema A. Murthy},
  title = {{Evidence of Task-Independent Person-Specific Signatures in EEG using Subspace Techniques}},
  journal = {IEEE Trans. Inf. Forensics Secur.},
  year = {2021},
  url = {https://openreview.net/forum?id=D4ii19pQHc},
}
Figure