A Sparse Representation Classification Framework for Person Identification and Verification Using Neurophysiological Signals

Abstract

Brain biometrics has received increasing attention from the scientific community due to its unique properties in comparison to traditional biometric methods. Many studies have shown that EEG features are distinct among individuals. SSVEP signals, generated by stationary localized sources and distributed sources in the parietal and occipital regions of the brain, serve as a reliable basis for biometrics. In this study, we present a novel approach that leverages the spatial patterns of brain responses elicited by visual stimulation at specific frequencies. Specifically, we propose integrating common spatial patterns with Sparse Representation Classification (SRC) frameworks for person identification and verification. The use of common spatial patterns enables the design of personalized spatial filters, which play a crucial role in constructing the dictionary used by SRC frameworks. We conducted extensive evaluations of the proposed method, comparing it with several traditional approaches using two SSVEP datasets. Our analysis also explored a broad range of flickering frequencies in the SSVEP experiments. The results from these datasets demonstrated the effectiveness of our approach for person identification and verification, achieving an average correct recognition rate above 90% across various visual stimulus frequencies and short durations of electrophysiological signals.

1. Introduction

Human–computer interactions have significantly transformed society, fostering greater connectivity among individuals and communities [1]. Recent advances in these interactions have increased interest in utilizing cutting-edge techniques to enhance intelligent systems, including optimization, machine learning, virtual reality (VR), augmented reality (AR), mixed reality (MR), interactive games, simulations, serious games, and emotion and mood analysis. Biometrics is the science of establishing an individual’s identity based on their physical and/or behavioral traits, either in a fully automated or semi-automated manner [2]. Conventional biometrics, such as fingerprints, face, iris, voice, and DNA, have been extensively studied in the literature and widely implemented in real-life applications. However, these conventional biometrics have weaknesses, as they are non-cancelable, disclosable, easily stolen, and lack liveness detection [2,3,4,5]. Brain electrical activity, measured at the scalp through electroencephalography (EEG), is a potential alternative to conventional biometrics, as it is influenced by genetic factors and exhibits high individuality among people. Recent studies in the biometric research community support this claim [6,7]. Compared with conventional biometrics, EEG possesses several unique advantages, such as being cancelable, non-disclosable, not easily stolen, and providing liveness detection [3].

The unique advantages of EEG-based biometrics—including cancelability, non-disclosability, and liveness detection—originate from the inherent properties of brain signals and their acquisition process. Unlike conventional biometrics, such as fingerprints or facial recognition, EEG signals are not physically exposed and cannot be easily copied, stolen, or forged, thus making them non-disclosable. Additionally, EEG patterns vary over time and can be reconfigured using different cognitive tasks or stimulus paradigms, allowing for cancelability, where a compromised biometric template can be reset by acquiring a new EEG-based profile. Liveness detection is an intrinsic advantage of EEG biometrics because signal acquisition requires active participation from the subject. Unlike facial images or fingerprints, which can be replicated using photographs or other artificial means, EEG signals are generated in real time, ensuring that the data originate from a live user. These advantages collectively make EEG-based biometrics inherently more secure and resistant to spoofing attacks.

Person verification and identification in EEG biometric systems share foundational similarities, as both rely on analyzing unique EEG signal characteristics to distinguish individuals [3,8,9]. Both approaches involve EEG data preprocessing, feature extraction, and classification to evaluate identity-related patterns. However, their objectives and methodologies differ in subtle yet significant ways [8,10,11]. Verification focuses on confirming a claimed identity through a one-to-one comparison, assessing whether the presented EEG data match the stored template of the claimed user. In contrast, identification seeks to determine the individual’s identity through a one-to-many comparison, matching EEG data against all stored templates in the database. This distinction leads to differences in system design: verification systems emphasize minimizing false acceptances (impostors accepted) and false rejections (genuine users rejected), while identification systems prioritize accurate recognition among multiple candidates, requiring robust scalability as the database grows. Additionally, identification systems must be able to handle the possibility of users not being in the database, a scenario not encountered in verification tasks. These distinctions, although subtle, require tailored approaches to feature selection, matching algorithms, and evaluation metrics for each task.

EEG-based biometric systems require careful consideration of acquisition protocols, feature engineering, and classification techniques [3]. Two types of EEG devices are commonly used: medical-grade systems, which feature high-quality wet sensors and numerous electrodes, and low-cost systems with fewer electrodes and dry sensors, which improve usability but reduce signal quality [3,12,13,14]. The selection of brain response types is also critical, with three main stimulation protocols: resting state (least demanding but prone to variability), sensory stimuli (requiring external devices but offering task-specific responses), and cognitive tasks (intentional mental activities that may lack consistency) [6,11,15,16,17,18,19,20]. Furthermore, feature engineering plays a crucial role in person identification, with commonly extracted features spanning time, frequency, and connectivity domains. Techniques such as autoregressive modeling, power spectral density (PSD), wavelet transforms, and brain connectivity measures—including phase locking values (PLV) and functional connectivity—have been widely applied, although spatial patterns remain underexplored [11,14,15,17,18,21,22]. Additionally, time-frequency (TF) analysis has been extensively used in EEG-based biometrics to address the non-stationary nature of EEG signals, enabling the extraction of meaningful features that capture both temporal and spectral variations [23,24].

Various classification methods have been employed, including similarity-based approaches like kNN, kernel-based methods such as SVM, linear discriminant analysis (LDA), and deep neural networks (DNNs) [3,5,11,12,20,22,25,26]. While kNN uses distance metrics to compare feature samples, SVM leverages the “kernel trick” to map data into high-dimensional spaces for easier classification. Additionally, LDA assumes normally distributed classes and projects data into lower-dimensional spaces, making it a popular choice. Deep learning, especially CNNs, has gained attention for using either raw EEG data or engineered features as inputs, showing promising results in person identification. However, deep networks require large datasets and significant training times, which can hinder their practical deployment in EEG biometric systems.

Although SSVEP signals are rarely used for person identification, they hold significant potential, particularly when adopting principles from BCI systems such as SSVEP-based spellers [27]. Most studies have focused on small datasets with limited stimulus frequencies and employed features such as Short-Time Fourier Transform, normalized variances, Mel-frequency cepstral coefficients (MFCCs), and autoregressive (AR) reflection coefficients [10,24,28,29]. These features have been paired with classifiers like kNN [24,29] or Convolutional Neural Networks (CNNs) applied to raw and enhanced SSVEP signals [10]. However, a major limitation is the small number of stimulus frequencies used, restricting design flexibility and the potential integration of biometric recognition with functionalities like password-based spellers. Furthermore, while most studies have focused on temporal and frequency features, the potential of spatial patterns in SSVEP signals remains unexplored. Spatial patterns of EEG have not been investigated extensively (to the best of the authors’ knowledge), besides [26], who used spatial filters in Auditory Evoked Potentials (AEP), and [27], who used spatial filters in SSVEP signals in conjunction with deep learning. Incorporating spatial filtering approaches, which have shown promise in other EEG modalities or tasks, could enhance the discriminative power of SSVEP-based biometric systems, presenting an important direction for future research.

Sparse Representation Classification (SRC) has proven to be highly effective in EEG signal processing, especially when spatial filters are used to construct an overcomplete dictionary [30,31,32,33]. Unlike traditional classification methods, SRC represents a test sample as a sparse linear combination of training samples [34]. This makes it particularly suitable for EEG signals, which are often noisy and non-stationary. By constructing overcomplete dictionaries and solving optimization problems to find the sparsest solution, SRC ensures that the representation aligns with the discriminative features of the correct class, even under challenging conditions like small training datasets or high variability between sessions [31,32]. Moreover, advanced SRC variants, such as those incorporating group sparsity or Bayesian frameworks, further enhance robustness by accounting for inter-class and intra-class structures in the data, enabling more accurate and generalizable classifications [31,32]. These strengths make SRC a valuable approach in EEG biometrics, especially for applications requiring high reliability in dynamic and noisy environments. While SRC has shown its usefulness in EEG processing, it has not been used in the context of EEG-based biometric systems. In this work, we propose using SRC for EEG-based biometric purposes. More specifically, the proposed SRC scheme is used for person identification and verification.

Our work introduces several key contributions to the field of EEG-based biometrics, particularly the use of SSVEP signals for person identification and verification. By leveraging the discriminative potential of SSVEP responses, we explore innovative methodologies that enhance the performance and applicability of biometric systems. A notable contribution is the integration of spatial patterns into the analysis, which enables a deeper understanding of individual EEG characteristics and improves system accuracy. Furthermore, we apply the Sparse Representation Classification (SRC) framework to both person identification and verification tasks, demonstrating its effectiveness in handling these machine learning tasks. Our study also addresses the challenges posed by using short-duration SSVEP responses, showing that, while certain methods, such as DFN, may struggle under these conditions, the proposed approaches exhibit robustness and adaptability. Lastly, we provide a comprehensive analysis of the differences between verification (authentication) and identification in EEG-based biometric systems, highlighting their distinct requirements and implications. These contributions collectively advance the understanding and development of EEG-based biometric technologies, paving the way for more efficient, accurate, and user-friendly systems.

The proposed framework presents an innovative combination of Sparse Representation Classification (SRC) and Filter-Bank Common Spatial Patterns (FBCSP), aiming to provide a more robust and efficient method for SSVEP-based person identification and verification. Unlike the study by [27], which relied heavily on deep learning techniques that require large datasets and high computational resources, our approach offers a more interpretable and lightweight solution that is suitable for real-time applications and limited-data scenarios. Furthermore, in contrast to [33], which applied SRC to the classification of mental states, our work specifically adapts SRC for biometric authentication, addressing the distinct challenges associated with recognizing individuals based on their EEG signals. The key novelty of our method lies in constructing an incoherent dictionary using FBCSP, which enhances the discriminative capacity of the SRC algorithm. This results in a framework that excels in both accuracy and robustness while performing well with short-duration EEG signals.

This paper is organized as follows. In Section 2, we present information about the SSVEP datasets and the proposed methodology for person identification and verification. More specifically, we describe the spatial filtering method and two SRC schemes for biometric purposes. After that, in Section 3, we provide information about the experimental settings of our experiments. Then, in Section 4, we present a comprehensive comparison of our approach with well-known classifiers for the same problem. Finally, in Section 5, we provide a discussion and concluding remarks related to our work and its future directions.

Table 1. Notations and descriptions.

Notation	Description
EEG	Electroencephalography
SSVEP	Steady-State Visual Evoked Potentials
SRC	Sparse Representation Classification
VR	virtual reality
AR	augmented reality
MR	mixed reality
PSD	power spectral density
PLV	phase locking values
TF	time frequency
SVM	Support Vector Machine
kNN	k-Nearest Neighborhood
LDA	Linear Discriminant Analysis
DNN	Deep Neural Network
CNN	Convolutional Neural Network
BCI	Brain–Computer Interface
MFCCs	Mel-frequency cepstral coefficients
AR	autoregressive
FBCSP	Filter-Bank Common Spatial Patterns
CCA	Canonical Correlation Analysis
TRCA	Task-Related Component Analysis
SCI	sparsity concentration index
CRR	correct recognition rate
Tr	duration of training set
Te	duration of test set
ROC	Receiver Operating Characteristic
TPR	true positive rate
FPR	false positive rate
TAR	true acceptance rate
FAR	false acceptance rate
DFN	Deep Feedforward Network
MLP	multilayer perceptron
TW	time window
PI	person identification
STFT	Short-Time Fourier Transform
SNR	signal-to-noise ratio

summarizes the main notations, along with descriptions, that are in this paper.

2. Materials and Methods

2.1. SSVEP Datasets

In this study, we utilized two benchmark SSVEP datasets for person identification (PI): the Speller dataset and the BETA dataset. Here, we provide a brief description of each dataset. The Speller dataset [35] contains SSVEP responses from 35 subjects, with 40 distinct stimuli. The stimulation frequencies range from 8 Hz to 15.8 Hz in 0.2 Hz increments, with a phase difference of $0.5\pi$ between adjacent frequencies. EEG signals were recorded using the SynAmps2 EEG system with 64 channels, following the extended 10–20 system. In this study, we selected nine channels covering the occipital and parietal-occipital regions (Pz, PO5, PO3, POz, PO4, PO6, O1, Oz, and O2). Each subject completed six blocks, with each block consisting of a 5 s visual stimulus presentation for each of the 40 targets. After extracting EEG trials, the signals were bandpass filtered between 7 and 90 Hz using an infinite impulse response (IIR) filter with MATLAB’s filtfilt function. The BETA dataset [36] follows a similar configuration as the Speller dataset but includes a larger sample size of 70 subjects. As in the Speller dataset, we used the same nine occipital and parietal-occipital channels (Pz, PO5, PO3, POz, PO4, PO6, O1, Oz, and O2). Each subject participated in four blocks, where the visual stimulus duration varied: 2 s for subjects S1–S15 and 3 s for subjects S16–S70. Further details on the experimental setup can be found in [36]. It is important to emphasize that the selection of these nine channels was based on the well-established fact that SSVEP responses are most prominent in the occipital and parietal-occipital regions, which are directly involved in visual processing. Thus, our approach assumes that each individual exhibits unique brain patterns in response to SSVEP stimuli, contributing to person identification.

An SSVEP dataset is a collection of multi-channel EEG trials ${\{{\mathbf{X}}_1^{\left(s\right)},{\mathbf{X}}_2^{\left(s\right)},\dots ,{\mathbf{X}}_M^{\left(s\right)}\}}_{s=1}^{N_s}$, where M is the number of trials of a participant, $\left(s\right)$ is the index of the participant, and $N_s$ is the number of participants (or classes). Each ${\mathbf{X}}_m^{\left(s\right)},m=1,\dots ,M,s=1,\dots ,N_s$ is a matrix of size $N_{ch}\times N_t$, where $N_{ch}$ is the number of channels and $N_t$ is the number of samples. Additionally, we assume that the multi-channel EEG signals are centralized since, in practice, the EEG trials are bandpass filtered or detrended. Furthermore, for purposes of PI, we assume that all SSVEP responses are acquired using visual stimuli that are flashing at the same frequency.

2.2. Spatial Filtering and Filter-Bank Common Spatial Patterns

Spatial filtering methods have found extensive application in EEG signal processing, ranging from the recognition of SSVEP responses [37] and motor imagery responses [38] to biometric applications [26]. Given the EEG trial $\mathbf{X}$, a spatial filter $\mathbf{w}:N_{ch}\times 1$ is the weights describing the linear combination of EEG channels with the overarching goal of producing a spatially filtered EEG trial (i.e., ${\mathbf{X}}^T\mathbf{w}$) with higher SNR than the original EEG trial (i.e., $\mathbf{X}$). The Filter-Bank Common Spatial Patterns (FBCSP) method [39] is a supervised learning technique for extracting discriminative features from EEG data by optimizing spatial filters across multiple frequency bands. It aims to maximize signal variance for one class while minimizing it for others, enabling effective separation of classes. The method generates features by projecting EEG signals onto spatial filters and calculating log-transformed variances. For multiclass problems, such as person identification, FBCSP employs a one-versus-rest strategy to compute individualized spatial filters, capturing unique spatial patterns for each individual across frequency bands. This approach combines efficiency and robustness, making it well suited for EEG-based biometric systems and other classification tasks requiring high discriminative power. For the extraction of FBCSP features, we follow the same procedure as that in [27]. The adaptation of FBCSP for person identification and verification lies in its ability to optimize spatial filters that enhance subject-specific discriminative features in EEG signals. Unlike its conventional application in motor imagery and BCI systems, where FBCSP is used to maximize variance between task-related EEG components, the proposed methodology modifies the standard FBCSP approach by tailoring the spatial filters to emphasize individual differences in SSVEP responses. This is achieved by applying FBCSP in a biometric-oriented fashion, where the learned spatial filters are designed to enhance subject-specific neurophysiological patterns rather than task-related components.

FBCSP is particularly well suited for SSVEP-based biometrics due to its ability to extract discriminative spatial features while effectively handling the multi-frequency nature of SSVEP responses. Compared to other spatial filtering techniques, FBCSP optimizes spatial filters to maximize variance differences between classes, leading to improved classification performance. Traditional methods such as Laplacian filtering primarily focus on enhancing the local signal-to-noise ratio (SNR) by reducing interference from neighboring electrodes. While effective for general EEG applications, Laplacian filtering does not explicitly optimize class separability, making it less ideal for biometric tasks where distinguishing between individuals is crucial. In contrast, FBCSP’s individualized spatial filters capture unique neurophysiological patterns that enhance subject recognition.

Other specialized SSVEP spatial filtering methods, such as Canonical Correlation Analysis (CCA) and Task-Related Component Analysis (TRCA) [37], are commonly used in SSVEP-based Brain–Computer Interfaces (BCIs). CCA identifies frequency components by maximizing correlations between EEG signals and predefined sine-cosine reference signals. However, it assumes stationary frequency responses across subjects, making it less suitable for biometrics, where individual variability is essential. TRCA enhances SSVEP detection by extracting temporally consistent components from multiple trials of the same stimulus but is not designed to model inter-subject differences effectively. Unlike CCA and TRCA, which are primarily signal detection methods, FBCSP learns spatial filters that are individualized for each subject, improving biometric recognition.

2.3. Sparse Representation Classification (SRC) Framework

SRC frameworks utilize training samples to form an overcomplete dictionary, enabling test samples to be represented as a sparse linear combination of class-specific training examples. In EEG studies, this implies that the brain features of a test sample can be approximated through a sparse linear combination of features from the same class. This subsection briefly outlines the SRC approach.

Given the labeled dataset $\mathcal{D}={\left\{({\mathbf{f}}_i,{\ell }_i)\right\}}_{i=1}^N$, where ${\mathbf{f}}_i$ represents the feature vectors and ${\ell }_i$ represents the corresponding labels, we can collect all the feature vectors in a matrix, $\mathbf{\Phi }\in {\Re }^{N_{ch}\times N}$. The core principle of SRC is that the label of the test vector $\mathbf{y}\in {\Re }^{N_{ch}}$ can be expressed as a linear combination of training samples from all classes with known labels:

$$\mathbf{y}=\mathbf{\Phi }\mathbf{\alpha }$$

(1)

where $\mathbf{\Phi }\in {\Re }^{N_{ch}\times N}$ is the dictionary matrix comprising all training vectors, N is the number of training vectors, and $\mathbf{\alpha }\in {\Re }^N$ is the coefficient vector. Furthermore, in the presence of noise, the model is extended as

$$\mathbf{y}=\mathbf{\Phi }\mathbf{\alpha }+\mathbf{e}$$

(2)

where $\mathbf{e}\in {\Re }^{N_{ch}}$ is the noise term with bound energy ${\parallel \mathbf{e}\parallel }_2\le ϵ$. In this case, the coefficients $\mathbf{\alpha }$ are found by solving the following minimization problem:

$$\widehat{\mathbf{w}}=arg\underset{\mathbf{\alpha }}{min}{\{\parallel \mathbf{y}-\mathbf{\Phi }\mathbf{\alpha }\parallel }_2^2+{\rho \parallel \mathbf{\alpha }\parallel }_1\}.$$

(3)

2.3.1. Sparse Representation Classification (SRC) for Identification

Having established that a test vector can be represented as a linear combination of training vectors, we now define a classification rule based on this representation. Specifically, classification is determined using the residuals of the linear combination. Let ${\delta }_c(·):{\mathbb{R}}^N\to {\mathbb{R}}^N$ be a function that selects the coefficients corresponding to class c. The residual for each class is then computed as $r_c\left(\mathbf{y}\right)={\parallel \mathbf{y}-\mathbf{\Phi }{\delta }_c\left(\widehat{\mathbf{\alpha }}\right)\parallel }_2,c=1,\dots ,C.$ The class of the test signal is assigned based on the minimum residual: $\mathrm{class}\left(\mathbf{y}\right)=arg{min}_c\left\{r_c\left(\mathbf{y}\right)\right\}.$ Thus, the classification process consists of two key steps: (1) solving the minimization problem to estimate $\widehat{\mathbf{\alpha }}$ and (2) applying the classification rule based on residual minimization.

Numerous extensions of the standard SRC algorithm have been introduced, aiming to enhance the optimization process in Step 1 and improve the residual computation in Step 2. In this study, we employ the approach proposed in [33], which exploits the manifold structure of the data by integrating a specialized prior based on graph theory while enforcing sparsity. A key element of this method is the training sample matrix $\mathbf{\Phi }$, or dictionary, which plays a pivotal role along with the solver responsible for computing the linear combination coefficients. The overall system architecture for person identification, depicted in Figure 1, comprises four primary components: EEG signal acquisition and preprocessing, spatial filter learning, FBCSP feature extraction, and dictionary construction. During the enrollment phase, multi-channel EEG signals are captured from users following the SSVEP protocol. These signals undergo spatial filtering to derive spatial filters and extract FBCSP features, which are subsequently used to build the dictionary. Both the spatial filters and extracted features are then stored in the database. In the authentication phase, newly acquired EEG signals are processed using the stored spatial filters to extract FBCSP features, after which residuals are computed to authenticate the user. Finally, using the fast marginal likelihood maximization procedure [40,41], it is possible to adaptively construct the dictionary and obtain a fast version of the proposed SRC algorithm, as shown in Algorithm 1.

Algorithm 1 SRC for identification

Require: Training samples, $\mathbf{\Phi }$, with their corresponding labels, ℓ, one test sample, $\mathbf{y}$, trade-off parameter $\lambda$, and number of nearest neighborhoods, k. Construct graph Laplacian matrix, L. Iterate over Equations (8), (9), (12) and (13), from [33], to find $\widehat{\mathbf{\alpha }}$ Calculate the residuals: $r_c\left(\mathbf{y}\right)={\parallel \mathbf{y}-\mathbf{\Phi }{\delta }_c\left(\widehat{\mathbf{\alpha }}\right)\parallel }_2$, $c=1,\dots ,C$ Ensure: $class\left(\mathbf{y}\right)=arg{min}_c\left\{r_c\left(\mathbf{y}\right)\right\}$

2.3.2. Sparse Representation Classification (SRC) for Verification

Sparse Representation Classification (SRC) for verification is fundamentally different from conventional classification tasks, as it operates more like a one-class classification or distribution learning problem [42,43]. In verification scenarios, the model only learns the characteristics of the legitimate user (the enrolled class) without access to impostor data during training. This makes SRC particularly suitable, as it focuses on representing the test sample as a sparse linear combination of the stored templates of the legitimate user. The absence of impostor data challenges typical discriminative models but aligns well with SRC’s generative nature, where the emphasis is on reconstructing the query sample using the enrolled user’s data. This approach enables effective verification, even in cases where impostor data are unavailable or impractical to collect, ensuring that the system relies solely on the uniqueness of the legitimate user’s EEG patterns.

To enhance its robustness in rejecting invalid test samples, SRC employs a dual strategy involving both residuals and the sparsity concentration index (SCI) [34]. While residuals quantify how well a test sample can be reconstructed from the training data, the SCI measures how concentrated the sparse representation is on a single class [34]. This combination allows SRC to effectively differentiate between genuine and impostor inputs, even when the impostor data closely resemble multiple classes in the training set. By separating identification and validation tasks, SRC avoids false positives that might arise from ambiguous inputs. This capability is particularly valuable in EEG biometric verification, where rejecting impostors and handling noisy, high-dimensional data are critical challenges. Together, these features make SRC a robust and flexible solution for biometric verification systems. Below, we provide Algorithm 2, based on SRC, for the verification task. The overall system architecture for person verification is illustrated in Figure 2. The main difference in this architecture is related to the decision-making step. In this task, the decision is binary and relates to whether a user belongs to our database.

Algorithm 2 SRC for verification

Require: Training samples, $\mathbf{\Phi }$, with their corresponding labels, ℓ, one test sample, $\mathbf{y}$, trade-off parameter $\lambda$, threshold r, and number of nearest neighborhoods, k. Construct graph Laplacian matrix, L. Iterate over Equations (8), (9), (12) and (13), from [33], to find $\widehat{\mathbf{\alpha }}$ Calculate concentration index: $SCI={\displaystyle \frac{C·{max}_c\parallel {\delta }_c\left(\widehat{\mathbf{\alpha }}\right)\parallel /\parallel \widehat{\mathbf{\alpha }}\parallel -1}{C-1}}$, $c=1,\dots ,C$ Ensure: if $SCI>r$ return Yes else return No

3. Experimental Design

3.1. Performance Metrics

Since the person identification task is a one-against-all classification task, we choose the correct recognition rate (CRR) as the performance metric. The CRR is a well-known performance metric that is extensively used for person identification purposes, and it is defined as the average over the diagonal elements of the resulting confusion matrix [18]. However, besides the conventional CRR metric, other factors affecting the overall usability of the system should be considered when assessing the practical usability of any reported EEG-based recognition system. In this direction, the usability measure proposed in [5] is defined as

$$U={\displaystyle \frac{N\times CRR}{Tr+K\times Te}}$$

(4)

where N is the number of subjects, K is the number of electrodes, $Tr$ is the duration of the training set, and $Te$ is the duration of the test set.

For the verification task, performance was evaluated by adopting the Receiver Operating Characteristic (ROC) curve. Note that in traditional ROC curves, the TPR vs. FPR is plotted, but in our case, we modify the axes to use the TAR and FAR. In biometric verification, the ROC curve is a powerful summary of the relationship between the TAR and FAR, showing how well the system balances user acceptance and impostor rejection.

3.2. Experimental Settings

Typically, in an SSVEP experiment, a number of experimental settings must be defined and reported. More specifically, it is important to report the following settings:

• $N_p$: the number of subjects;
• $N_B$: the number of blocks;
• $N_f$: the number of stimulus frequencies;
• $N_{ch}$: the number of EEG channels;
• $D_{tr}$: the trial duration.

For the two SSVEP datasets, these settings are provided in

Table 2. Experimental settings for the two SSVEP datasets.

Parameter	Speller	BETA
$N_p$	35	70
$N_B$	6	4
$N_f$	40	40
$N_{ch}$	9	9
$D_{tr}$	5 s	2 s

.

Two sets of experiments were designed to simulate real-world EEG-based biometric scenarios: identification and verification. The identification scenario involved determining the identity of a user from a database of enrolled individuals by comparing their EEG data to stored templates. In contrast, the verification scenario focused on validating whether a user is who they claim to be through a one-to-one comparison with their specific enrolled template. By designing experiments that reflect both scenarios, our study ensures that the proposed SRC method is thoroughly evaluated under realistic conditions, addressing the unique challenges of each use case. This approach not only highlights the adaptability and effectiveness of the proposed SRC method but also provides insights into its potential deployment in diverse real-world applications, where identification and verification tasks are crucial components of EEG-based biometric systems.

Furthermore, to train the reported classification schemes in the case of the identification scenario, we adopted the leave-one-block-out (LOBO) cross-validation approach, where the block contained an SSVEP trial from each subject (total $N_p$ trials). For the identification task, we compared our method with the following methods:

• SVM-FBCSP: The SVM classifier with a linear kernel, where FBCSP features are used to perform person identification;
• kNN-FBCSP: The kNN classifier with FBCSP features;
• Deep Feedforward Network (DFN) [27]: A deep neural network architecture based on a multilayer perceptron (MLP), specifically designed to use SSVEP responses for human recognition.

A different cross-validation approach was used for the verification scenario. The verification scenario was applied to the BETA dataset due to the large number of subjects in this dataset. More specifically, the first three blocks from subjects 1–35 were used in the enrollment phase, while the first block from subjects 36–70 (impostors) and the fourth block from subjects 1 to 35 were used in the verification phase. It is important to note that the testing dataset contained impostors, as well as subjects who participated in the enrollment phase. Also, in this scenario, we compared our method with the one-class SVM method [42,43].

The overall data analysis procedure consisted of various steps. First, the EEG data were temporally filtered; then, the spatial filters were learned to extract the FBCSP features; and finally, the FBCSP features were used as input to the classifier or to create the overcomplete dictionary in the case of the SRC algorithm. Additionally, for the Speller and BETA datasets, we used the nine channels from the occipital and parietal-occipital areas (Pz, PO5, PO3, POz, PO4, PO6, O1, Oz, and O2).

4. Results

In this section, we provide a detailed presentation and analysis of the results obtained from our experiments. Our goal is to evaluate the performance of the proposed method under various conditions and scenarios, including both identification and verification tasks. Additionally, we compare the outcomes with baseline methods to highlight the advantages and limitations of the methods. Key metrics, such as the CRR, ROC curves, and usability, are discussed, offering insights into the practical implications of the findings. The results aim to provide a comprehensive understanding of the proposed SRC method’s capabilities and its applicability in real-world EEG-based biometric systems.

4.1. Identification Results

SSVEP signals are generated by stationary localized sources and distributed sources in the parietal and occipital areas of the brain. Our overarching goal in these experiments was to determine whether SSVEP signals contain information that can be used to discriminate individuals by using the proposed SRC scheme. Furthermore, we determined whether this information contained distinct spatial patterns among individuals. In Figure 3, we present the results of our experiments on the two datasets. Our intention here was to examine whether our algorithm could identify persons using SSVEP signals from various frequencies of the stimulus. In this set of experiments, the length of an SSVEP trial was 5 s for the Speller dataset and 2 s for the BETA dataset. In Figure 3, we can observe that the SRC method achieved the best performance across both datasets. Additionally, we examined the performance for each frequency (see Figure 4). In this figure, we can observe that in most stimuli (i.e., frequencies), the SRC method achieved the best performance. Overall, the proposed SRC method achieved better person recognition than all other methods. Importantly, this observation could be extended to a wide range of stimuli.

To determine whether the observed differences in Figure 3 and Figure 4 were statistically significant, we conducted one-way ANOVA to compare the effect of classification methods on the CRR values. For the Speller dataset, we found that there was a significant difference in accuracy among the classification methods at the p < 0.05 level for the four methods F(3,156) = 328.27, p< 0.001. Furthermore, post hoc analysis revealed that the proposed method had a significantly different accuracy from the SVM and kNN methods but not from the DFN, even though the SRC method had a higher CRR than the DFN. For the BETA dataset, we found that there was a significant difference in accuracy among the classification methods at the p< 0.05 level for the four methods F(3,156) = 237.81, p< 0.001. Furthermore, post hoc analysis revealed that the proposed method had a significantly different accuracy from all other methods. These results indicate that the proposed SRC method had a statistically significant effect on the final model’s performance.

In general, it is desirable to use SSVEP signals with a short duration because they contribute to the design of EEG-based biometric systems that are more comfortable and provide a better user experience. Hence, we performed additional experiments to study the performance of methods with different SSVEP signal durations. In

Table 3. CRRs on the Speller and BETA datasets with respect to the TW.

Methods
TW	SRC	DFN	SVM	kNN
Speller
1 s	94.38	92.57	85.71	85.24
2 s	97.81	97.52	91.33	90.86
3 s	98.48	98.76	92.76	92.95
4 s	98.95	98.19	94.38	93.05
5 s	99.31	99.01	94.77	93.27
BETA
1 s	84.29	79.79	76.79	81.43
2 s	92.69	89.54	81.87	87.31

Note: The provided CRR (%) values are averaged over all frequencies for various TWs.

, we present the averaged CRR over all stimulus frequencies for various SSVEP trial durations on both datasets. We can observe that the SRC method achieved the best recognition rate among all methods across all trial durations, except in the case where the TW = 3 on the Speller dataset. These findings demonstrate that the SRC method consistently outperformed other methods across different trial durations, demonstrating its robustness and adaptability to varying experimental conditions.

In our study, we also evaluated the usability of the comparative methods to obtain a more comprehensive view of the proposed algorithms and to provide comparisons with other EEG-based PI systems reported in the literature. The usability results, as presented in

Table 4. Usability scores on the Speller and BETA datasets with respect to the TW.

Methods
TW	SRC	DFN	SVM	kNN
Speller
1	97.16	95.29	88.23	87.75
2	79.61	79.38	74.34	73.96
3	66.28	66.47	62.43	62.56
4	56.77	56.34	54.15	53.39
5	49.66	49.51	47.38	46.63
BETA
1	393.35	372.35	358.35	380.01
2	270.35	261.16	238.79	254.65

Note: The provided CRR (%) values are averaged over all frequencies for various TWs.

, reveal clear distinctions in the performance of the evaluated methods—SRC, DFN, SVM, and kNN—across different trial durations on the Speller and BETA datasets. The SRC method consistently achieved the highest usability scores, demonstrating its superior efficiency and adaptability. On the Speller dataset, the SRC method outperformed the other methods across all trial durations, with a peak usability value of 97.16 at a trial duration of 1 s, highlighting its ability to maintain high performance at minimal trial durations. Similar observations were made with respect to the BETA dataset, where the SRC method achieved a usability score of 393.35 at a trial duration of 1 s, significantly outperforming the other methods. Overall, the results demonstrate that the SRC method is the most effective method across both datasets and all trial durations besides one, showcasing its suitability for applications where high usability and short EEG trials are essential. The DFN proved to be a reliable alternative, especially for scenarios involving longer trial durations. In contrast, the SVM and kNN methods yielded comparatively lower usability scores, indicating that they may be less optimal for high-performance EEG-based biometric systems. Finally, we observed that our approach outperformed those reported in [5] in terms of the usability metric. A key factor contributing to this superior performance was that our method achieved a high CRR even for very short EEG trial durations. While other methods also achieved high CRRs, they typically required longer EEG trials and a larger number of channels.

4.2. Verification Results

A practical biometric system should be able to correctly identify registered users and appropriately reject unregistered subjects. Our overarching goal in these experiments was to assess the ability of the proposed SRC scheme in verification tasks. In these experiments, we studied how well the proposed SRC algorithm (Algorithm 2) balanced user acceptance and impostor rejection. In Figure 5, the ROC curves of the proposed SRC method and the one-class SVM method are depicted, illustrating the comparative performance of these two algorithms in a verification task. The results clearly demonstrate that the SRC method (Algorithm 2) outperformed the one-class SVM method across all operating points. The Area Under the Curve (AUC) for the SRC method was 0.91, which was significantly higher than the 0.66 achieved by the SVM variant. This indicates that the SRC method achieved more robust and reliable performance in distinguishing between genuine and impostor samples. This suggests that the SRC method is better at correctly identifying genuine users while minimizing the acceptance of impostors, which is critical for biometric verification systems. The improved performance of the SRC method can be attributed to its ability to leverage sparse representations, which effectively capture the discriminative features of EEG signals, even in challenging scenarios with high variability or limited data. In contrast, the one-class SVM method struggled to match this level of discrimination, likely due to its reliance on simpler decision boundaries that may not fully exploit the complexity of EEG data.

5. Discussion and Conclusions

The analysis highlights SRC as the most robust and efficient algorithm across all scenarios, excelling particularly in handling short trial durations. It consistently achieves high CRRs and usability scores, demonstrating its ability to extract meaningful information from limited EEG data. Even under the challenging conditions of the BETA dataset, with its larger subject pool and shorter trial durations, SRC maintains superior performance, showcasing its adaptability and resilience to variability. Furthermore, the usability scores emphasize SRC’s efficiency, as it balances high recognition rates with minimal time and resource requirements. This efficiency makes SRC exceptionally well suited for real-world EEG biometric systems, where both accuracy and usability are critical. Its ability to consistently deliver robust results across diverse datasets and conditions underscores its potential as a reliable and scalable solution for EEG-based biometric tasks.

However, due to the limitation of available datasets and the challenges associated with data collection, the analysis in this study was conducted using EEG signals from single sessions. While single-session data provide valuable insights, they do not fully address the critical issue of permanence in EEG-based biometric systems, which requires evaluating the stability of EEG features across multiple sessions and over extended periods. Further studies leveraging multi-session EEG data are essential for gaining a deeper understanding of how neurophysiological changes, experimental conditions, and other factors impact system performance [11,23,27,44]. In future work, we aim to address these gaps by extending our investigation to multi-session settings, focusing on long-term data collection to evaluate the consistency and reliability of EEG features over time. This approach will provide a more comprehensive evaluation of the practicality and robustness of EEG-based biometric systems in real-world scenarios.

Inter-subject variability is a well-known challenge in EEG signal processing in general and in EEG-based biometric systems in particular, as individual differences in neurophysiological responses can affect the consistency and generalizability of classification metrics. In this study, the impact of inter-subject variability is partially mitigated by employing FBCSP, which optimizes spatial filters to capture subject-specific features, and SRC, which constructs an overcomplete dictionary that inherently adapts to individual differences. However, variations in factors such as cortical anatomy, cognitive state, and, particularly, session-to-session differences may still affect performance. While the current results demonstrate high accuracy across subjects, future work could incorporate domain adaptation techniques and subject-independent feature learning to further improve generalizability. Additionally, evaluating this methodology on larger and more diverse datasets would provide deeper insights into how inter-subject variability influences biometric recognition performance.

The proposed SRC-based biometric framework has the potential for further improvement by incorporating TF analysis, enabling a more comprehensive characterization of EEG signals. In this enhanced methodology, FBCSP could be initially employed to extract spatially filtered EEG signals. Subsequently, TF methods such as wavelet transforms and Short-Time Fourier Transform (STFT) could be applied, providing a joint representation that captures both the temporal and spectral characteristics of the signals. These techniques would effectively reveal dynamic variations in EEG signals, contributing to more informative feature extraction. The resulting TF-based features, including wavelet coefficients and frequency energy distributions, would add discriminative power and increase the robustness of biometric recognition. The final step would involve applying SRC to classify individuals by constructing an overcomplete dictionary that represents EEG responses as a sparse linear combination of training samples. Integrating TF methods into the framework would improve its ability to manage non-stationary EEG signals, enhance subject separability, and boost the overall accuracy and adaptability of the system for EEG-based person identification and verification.

The SRC method aligns well with several key learning paradigms, including self-representation learning [45], semi-supervised learning, lifelong learning, and few-shot learning. In the following paragraphs, we discuss how SRC could be extended to these learning paradigms in the context of EEG-based biometrics. First, the SRC method is inherently rooted in self-representation learning, and its framework could be further enhanced by explicitly leveraging the principles of this paradigm for EEG biometrics. In SRC, a test EEG sample is expressed as a sparse linear combination of training EEG samples, capturing the relationships and dependencies within the dataset. To extend this, self-representation regularization could be applied to ensure that the learned sparse coefficients reflect meaningful relationships between subjects, emphasizing intra-class coherence while reducing inter-class overlap. Furthermore, incorporating graph-based self-representation could model the inherent structure of EEG data, enabling more robust representations by leveraging both spatial and temporal correlations in the signals. Additionally, the dictionary could be dynamically updated to better reflect the evolving data distributions, ensuring that self-representation remains adaptive and robust over time. These enhancements to SRC, grounded in self-representation learning, could allow the framework to better exploit the unique characteristics of EEG signals for accurate and reliable person identification and verification.

Additionally, SRC could be extended to semi-supervised learning by leveraging both labeled and unlabeled data to enhance its performance, particularly in scenarios where labeled EEG data are limited [46]. One potential extension could be to incorporate unlabeled data into the dictionary, allowing the model to capture the underlying structure and manifold of the entire dataset rather than just the labeled samples. Graph-based regularization techniques could also be applied to encourage consistency between labeled and unlabeled data by propagating label information through a graph representation of the dataset. Additionally, pseudo-labeling strategies could be used to assign tentative labels to unlabeled EEG samples based on reconstruction errors or class proximity, enabling iterative refinement of the model. To address the unique challenges of EEG biometrics, adaptive constraints could be introduced to account for the non-stationarity and variability of EEG signals, ensuring that the learned representations remain robust across sessions. By integrating these semi-supervised learning principles, SRC could effectively utilize the wealth of unlabeled EEG data, improving its generalization and robustness for biometric applications.

SRC could also be extended to lifelong learning by incorporating dynamic mechanisms that allow the system to adapt to new data while preserving previously learned information. One potential extension could involve dynamic dictionary updates [33,41,47], where the dictionary is incrementally expanded to include new users or evolving EEG patterns, using online sparse coding or incremental learning techniques to ensure scalability. To address the time-varying nature of EEG signals, sliding-window approaches or adaptive weighting mechanisms could be used to emphasize recent data while maintaining historical context. Furthermore, domain adaptation techniques [48,49] could align feature distributions across sessions, mitigating the non-stationarity caused by neurophysiological changes or experimental variations. By implementing these extensions, SRC could become a robust lifelong learning framework capable of handling dynamic, real-world EEG biometric systems that require continuous adaptation and scalability.

Finally, SRC aligns naturally with the principles of few-shot learning [45,50] and could be extended to better address the challenges posed by limited EEG data in biometric applications [27]. In SRC, a test EEG sample is represented as a sparse linear combination of training samples, which inherently supports few-shot scenarios by focusing on the most relevant and class-specific data points. To enhance SRC for few-shot learning, prototype-based dictionaries could be introduced, where the dictionary would be constructed using representative prototypes of each class rather than full datasets, thereby reducing noise and computational complexity. Additionally, meta-learning techniques could be integrated to optimize the dictionary initialization or sparse reconstruction process for rapid adaptation to new subjects with minimal data. Class-specific regularization could further improve performance by emphasizing separability between classes even when only a few labeled samples are available. By leveraging these extensions, SRC could provide robust and scalable solutions for EEG-based person identification and verification, ensuring high accuracy even with minimal training data.