TASK-SPECIFIC BENCHMARKING OF EEG-BASED

BIOMETRIC ACROSS MULTIPLE TIME WINDOWS

(Submitted on IJCB CONFERENCE)

by Krishanu Deb and Deep Haloi

 WHAT IS EEG , WHY EEG AS BIOMETRICS ?

Electroencephalography (EEG) is a non-invasive technique that measures the electrical

activity of the brain. It involves placing electrodes on the scalp to detect the brain’s natural
electrical signals, typically in the form of brainwaves (like alpha, beta, theta, and delta
waves). The brain’s neurons communicate via electrical impulses, generating oscillatory
activity, which is captured by electrodes placed on the scalp. The signals are recorded as
waveforms representing brain activity across different frequency bands.
EEG signals are dynamic and individual-specific, making
them a strong modality for biometric au thentication. It is
used as a biometric because brainwave patterns are unique
to each individual, making EEG a reliable identifier. Unlike
more common biometrics such as fingerprints or facial
recognition, EEG signals are generated internally by the
brain, making them extremely difficult to fake or replicate.
 ABSTRACT
This study benchmarks various task-induced EEGstates—four cognitive tasks, two relaxed conditions, and an
emotion-induction task—across three public datasets: SAM40, FACED, and M3CV. To our knowledge, this is the
first comprehensive eval uation of EEG-based authentication across such diverse tasks and datasets.

We assess traditional features (time, frequency, time-frequency) and deep learning models like EEGNet across
multiple time windows: 0.5s, 1s, 2.5s, and 4s. Results show that arithmetic, Oddball, and emotion tasks yield the
most discriminative EEG patterns. Time-frequency features with logistic regression consistently outperform
others, achieving perfect accuracy and F1 scores also offering valuable guidance on task and window selection
for robust, interpretable EEG-based authentication systems

SHAP analysis highlights key STFT bins and CWT scales, enhancing interpretability.An additional approach is
employed by utilizing EEGNet, Following feature extraction by EEGNet, a 1D Convolutional Neural Network (1D
CNN) is applied as a downstream classifier to improve the model’s classification performance and capture
complex, hierarchical patterns in the EEG signals.

The results of this study demonstrate the potential of EEG-based classification systems in dis tinguishing between
subjects with high accuracy, even with datasets of varying sizes. This work establishes a foundation for further
exploration into the application of EEG-based systems in biomet ric authentication. The findings underscore the
viability of EEG as a secure, non-intrusive modality for personal identification, paving the way for its future
integration into authentication systems requiring high levels of security and personalization
 EEG DATASET OVERVIEW: STRUCTURE, PREPROCESSING, TIME
WINDOW SEGMENTATION, AND FEATURE EXTRACTION

The datasets used in this study—SAM40, M3CV, and FACED—vary in terms of participant count, task
diversity, and recording configurations but were standardized during the data preparation phase for
consistent analysis.

• The SAM40 dataset comprises EEG data from 40 subjects performing 4 cognitive tasks
(Arithmetic, Stroop, Mirror Image, and Relax), each recorded using 32 EEG channels at a
sampling rate of 128 Hz

• The M3CV dataset includes EEG recordings from 95 subjects across 3 tasks (Eye Close, Eye
Open, and P300), recorded with 65 channels at a higher sampling rate of 250 Hz. Each task was
divided into 30 trials of 4 seconds each, resulting in 90 trials per subject, with 1000 samples per
trial.

• The FACED dataset features data from 123 subjects who watched 28 emotion-eliciting video
clips, segmented into 168 trials of 5 seconds each, captured at 250 Hz across 32 channels,
resulting in 1250 samples per trial.
 To further enhance the analysis and enable temporal resolution studies, each of the
segmented trials across all datasets was additionally clipped into smaller windows of 0.5
seconds, 1 second, 2.5 seconds, and 4 seconds. This multiscale temporal segmentation not
only increased the number of training instances but also facilitated a detailed investigation
into how classification performance varies with respect to time window size.

To effectively analyze EEG signals, we extracted features from three domains: temporal, frequency,
and time-frequency.
• Temporal Features capture signal behavior over time and include statistical measures such as
mean, standard deviation, skewness, kurtosis, and Hjorth parameters (activity, mobility,
complexity). These features help in identifying amplitude variations and signal morphology.
• Frequency Features are derived using Fast Fourier Transform (FFT) and Power Spectral Density
(PSD). We calculated band power and relative power in standard EEG bands: delta (0.5–4 Hz),
theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (30–50 Hz). These features reveal
cognitive and emotional state information encoded in brain rhythms.
• Time-Frequency Features were extracted using Short-Time Fourier Transform (STFT) and Discrete
Wavelet Transform (DWT) to capture how frequency content evolves over time. From these, we
derived wavelet coefficients, energy, and entropy across time and scale, which are crucial for
analyzing the non-stationary nature of EEG signals.
 THE PROJECT PIPELINE

We implemented several traditional classification algorithms on the extracted features to evaluate baseline performance:
• Logistic Regression (LR): A simple and interpretable linear model useful for binary and multi-class classification.
• K-Nearest Neighbors (KNN): A non-parametric method that classifies samples based on majority voting from the nearest neighbors.
• Support Vector Machine (SVM): Effective in high-dimensional spaces and uses kernels for non-linear decision boundaries.
• Random Forest (RF): An ensemble of decision trees that improves accuracy and robustness.
• XGBoost (XG): A powerful gradient-boosting framework optimized for performance and scalability.
These models were trained using the temporal, frequency, and time-frequency features extracted from EEG clips.
 EEGNET
->

EEGNET+ 1DCNN
->
 ACCURACY, PRECISION , RECALL , F1 SCORE
RESULTS

Time Domain results Frequency Domain results

Time - Frequency Domain results

Deep Learning-based results

CONFUSION MATRIX AND ROC CURVE
(BEST)
 EXPLAINABLE AI AND SHAP ANALYSIS

Explainable AI (XAI)
In critical domains like Brain-Computer Interfaces (BCI), explainability is just as important as
accuracy. Explainable AI (XAI) aims to make machine learning models transparent and
understandable by providing insights into why a model makes a certain decision. This is essential
when working with complex models, especially in healthcare and neuroscience, where
understanding which brain regions or frequency bands influence classification can lead to more
interpretable and trustworthy systems.

SHAP Analysis
SHAP (SHapley Additive exPlanations) is a powerful and widely-used XAI method that explains
individual model predictions by computing Shapley values from cooperative game theory. These
values represent the contribution of each input feature to the final prediction.
Key benefits of SHAP:
• Provides local explanations (for individual predictions) and global explanations (overall feature
importance).
• Works with both traditional ML models (like XGBoost, SVM, etc.) and neural networks.
• Visualizes the positive or negative impact of each feature on the model output.
 WHY SHAP ANALYSIS IS REQUIRED HERE ??

In EEG classification tasks:

• We extract a large number of features from temporal, frequency, and time-frequency domains.
• Many of these features may not equally influence the model’s decision.
• SHAP helps identify the most important EEG features, such as:
⚬ Which EEG channels are most informative?
⚬ Which frequency bands (e.g., Alpha or Beta) drive specific cognitive/emotional states?
⚬ How much do temporal patterns contribute to classification?

By applying SHAP, we:

• Increase the trustworthiness and interpretability of the model.
• Gain neuroscientific insights into which brain activities correlate with the tasks or emotions.
• Enable clinical relevance, as practitioners can see which signals lead to specific diagnoses or
mental states.
 SHAP PLOT

The best-performing model was analyzed using SHAP (SHapley

Additive exPlanations) to improve interpretability and understand
model behavior. Specifically, the Logistic Regression classifier
trained on time–frequency features for the Oddball task (2.5-
second window) was examined. The SHAP summary plot in Figure
8.1 highlights the top 20 influential features, with STFT bins 12
and 13 showing the greatest impact, followed by bins 4, 3, 14, 7,
and 11. CWT features at scales 8–12, especially 9 and 10, also
contributed strongly, indicating that mid-range wavelet
resolutions effectively capture task-relevant dynamics. The SHAP
value distribution, skewed positively (0 to 0.8), suggests that the
presence of certain spectral patterns is more informative than
their absence. This analysis underscores the Oddball task’s
robustness and the importance of high-frequency spectral and
mid-scale wavelet features in EEG-based biometric classification.
 CONCLUSIONS

In this study, we introduced TaskBio-EEG to explore how different tasks, feature types, and models
affect EEG-based biometric authentication. We examined seven brain states triggered by tasks
using three public datasets. Our results show that tasks like arithmetic, Stroop color-word, and
oddball create clearer and more consistent EEG patterns than those related to relaxation or
emotions. We discovered that using time-frequency features with simple models like logistic
regression provides better outcomes compared to complex deep learning models. This method
balances performance, clarity, and 49 efficiency. Additionally, SHAP-based analysis helps identify
important brain patterns for distinguishing individuals, which boosts model reliability. This
benchmark serves as a valuable guide for future EEG-based biometrics research. It aids in selecting
appropriate tasks and promotes practical, easy-to understand modeling strategies for real-world
use. Moving forward, we aim to enhance this study by incorporating transfer learning, improving
performance across datasets, and using neural architectures that adapt to tasks, making biometric
systems more robust for everyday use. Future work will also explore explicit cross-dataset
evaluation and domain adaptation methods to ensure robust deployment across acquisition settings
THANK YOU
Any Questions ??