Among the different techniques used for measuring brain activity, Electroencephalography (EEG) is a method used for measuring voltage fluctuation inside the electrical field generated by a subject's brain. The sampling process involves placing surface electrodes onto a subject's scalp, capturing brain activity at different regions. The resulting EEG signals are non-stationary, meaning frequency values along time are not constant but variable, making them unpredictable. Recent advances in Deep Learning (DL) (Scarselli et al., 2008; Goodfellow et al., 2014; Vaswani et al., 2017) have allowed for new and improved applications of EEG signals. These applications include but are not limited to: EEG-based biometric systems (DelPozo-Banos et al., 2015), Brain-Computer Interfaces (BCIs) (Vaid et al., 2015), emotion recognition systems (Jenke et al., 2014), alertness analysis (Subasi, 2005) and medical diagnosis and progression assessment for neurological diseases such as Alzheimer's (Cassani et al., 2018) or epilepsy (Acharya et al., 2015). These systems are often composed of an automatic feature extraction layer that extracts representative features from the raw EEG signal, and a classifier for fitting a specific target feature. For example, Wilaiprasitporn et al. (2019) used a combination of Convolutional Neural Networks along with different type of temporal layers to learn distinctive features, and a fully connected layer for the purpose of classification. Similarly, Ullah et al. (2018) used a model composed of several 1-D convolutional layers for feature extraction and denoising to obtain representative features subsequently handled by a fully connected layer with a majority voting scheme.

In detail, this research focuses on biometrics, a process consisting of measuring and analyzing unique physical characteristics from humans to authenticate their unique identity. Among the methods employed for such a task, EEG has proven to be a robust biometrics method (Jayarathne et al., 2017). As pointed by Campisi and La Rocca (2014), EEG poses several advantages when compared to traditional exposed biometric modalities such as fingerprint or iris scanning. Namely: The advantage of being more robust against spoofing attacks (Revett, 2012), given the technological complexity of synthetically generating brain signals replicating the response for a specific individual. The advantage of being universal, meaning the sampling process is valid for anyone with no brain pathological conditions. Lastly, EEG poses the advantage of the aliveness detection problem not being present, meaning we can assume that the user is alive to produce brain signals sensor can read. This assumption does not necessarily hold in other biometric methods. There are two main applications for EEG in biometric systems: Person identification (PI), identifying an individual from a group of known subjects, and person authentication (PA), accepting or denying the identity of a particular individual. Both tasks' objective is to extract features with discriminating properties from EEG signals that a statistical model can later classify. Depending on the task subjects perform while their EEG recordings are taking place, it is possible to divide biometric systems into two groups, namely task-dependent (Wilaiprasitporn et al., 2019) and task-independent (Kong et al., 2018). Training a task-dependent system often means presenting the model data where all subjects perform a particular task. Whereas task-independent systems use data sampled from various tasks, making them more robust at handling unseen tasks at inference time. This work is centered around using EEG signals applied to task-dependent PI and is devoted to answering the following research question:

The structure of this document is as follows: Section 2 provides a brief literature review on the different methods used for EEG-Based Person Identification. Section 3 explains the design and methodology, with reproducibility and replicability in mind. Section 4 discusses the results and findings and finally Section 5 summarizes these informing possible future directions.

There exists evidence suggesting that different subjects produce different EEG responses (Marcel and Millán, 2007), with studies dating as back as the 1970's suggesting the uniqueness in EEG responses depends on each subject genetic information (Vogel, 1970; Anokhin et al., 1992). This unique property from EEG data makes it suitable for biometrics applications. However, building such applications has proven challenging to build and deploy given the non-stationary nature of EEG signals and the high quantity of noise generated during recording by, for example, muscle movements, eye blinking, electrode displacement. Additionally, there has been evidence showing that the EEG responses from an individual may vary depending on their emotions. For example, EEG responses may vary if the participant is emotionally attached to the watched person (Koelstra et al., 2009). EEG analysis often requires the extraction of high-level features fed into statistical models that automatically learn to find hidden patterns within the features. There are several different feature extraction techniques present in the literature. The feature extraction methods that are applied most often to this domain are Power Spectral Density (PSD) (Riera et al., 2007), Auto-Regressive (AR) model coefficients (Mohammadi et al., 2006; Zivot and Wang, 2006), Discrete Wavelet transform (DWT) (Guo et al., 2009; Murugappan et al., 2009), and Hilbert-Huang Transform (HHT) (Li et al., 2009). Among these techniques, wavelet-based feature extraction stands up given its performance at characterizing non-stationary data.

The available literature on person identification from EEG features dates as far as the 90's [see (Poulos et al., 1999a,b,c)]. Scholars used sets of features obtained from an AR model and a set of different Machine Learning (ML) classifiers obtaining accuracy scores ranging from 80 to 100% at classifying unique individuals from a pool of four total subjects. Mohammadi et al. (2006) trained a competitive neural network model with features obtained from an AR. In this case, scholars performed experiments training models for single-channel and multi-channel settings with up to three channels. They collected 24 s of EEG readings for 10 participants and used 15 s to train a model and a total of 24 s for testing. Results demonstrated that it was possible to obtain accuracies ranging from 80% up to 100% by using this technique on a larger pool of subjects. Brigham and Kumar (2010) used coefficients obtained for each channel from an AR model along with a Support Vector Machine (SVM) for PI. The scholars obtained testing classification accuracies close to 99% at identifying subjects from a pool of 120 subjects, suggesting that it is possible to obtain separable features from each individual. Shedeed (2011) used a combination of features obtained from applying the Discrete Fourier Transform and the Discrete Wavelet Transform followed by a voting scheme to retain the most discriminating features, which are input to a multi-layer perceptron classifier. However, their dataset was composed solely of three subjects, casting doubt on the generalization ability of their model. Thomas and Vinod (2018) used PSD features obtained from a single channel from the gamma (γ, 30–50 Hz) band along with simple correlation-based matching operations on a dataset composed of 109 subjects with an error rate of 0.0196. These results suggest that it is possible to apply a non-parametric operation over the raw EEG signal that maps input data onto a set of discriminative features for each subject.

Recent advances in DL have allowed the development of models with more expressing capabilities. Unlike approaches using hand-crafted features, DL models are often composed of two parts: a parametric feature extraction block and a classifier. The feature extraction block aims to learn representative features from the raw input signal automatically. The classifier uses the extracted features as input and predicts a target value such as a specific subject identity. Automatic feature extraction layers use spatial layers to reduce the dimensionality of the input data while trying to maintain informative features. This process usually consists of employing a combination of convolutional and pooling layers (Das et al., 2017, 2018). The input dimensionality of the data is usually not very large, given that there is a limited number of electrodes. The small input size leads to convolutional models with a relatively small number of layers, often ranging from 1 to 5 (Maiorana, 2020). Özdenizci et al. (2019) used a dataset composed of 128 subjects with readings obtained from different sessions and proposed an adversarial network from an invariant representation-learning perspective. Their goal was to create features that would be invariant between different recording sessions. Their approach consisted of an encoder convolutional layer to encode the raw signals and two classification layers. One of these layers is the identifier that predicts the correct person ID. The other one is the adversary layer that tries to predict the recording session ID. The assumption was that the encoder obtained features that the identifier could analyse but the adversary network could not, which led to session-invariant features. Additionally to the spatial layers, some approaches consider temporal layers to include temporal information in the features. Recurrent Neural Networks often include a combination of spatial and temporal layers (Chen et al., 2020). Das et al. (2019) proposed a model based on a combination of spatial layers (Convolutional) and temporal layers by employing Long Short Term Memory (LSTM). Results pointed to accuracies close to 100% on a pool of 109 subjects using the raw data from recordings of 64 channels. Wilaiprasitporn et al. (2019) proposed an alternative method using Gated Recurrent Units instead of the LSTM, achieving 99.9–100% test accuracy on the DEAP dataset (Koelstra et al., 2011) using all available channels. They also achieved 99.17% test accuracy, reducing the number of channels from 32 to 5, using their Frontal and Parietal (F3, F4, Fz, F7, and F8) configuration. Chen et al. (2020) demonstrated that it was possible to train models that automatically learn feature extraction techniques from the raw signal using spatial and temporal layers. These rich features are then injected into a classifier, where the final target feature is the subject ID. They achieved 96% test accuracy on a pool of 157 subjects sampled from four different experiments, demonstrating their model's ability to generalize.

The recent surge in interest in Graph Neural Network (GNN) architectures has also affected EEG-based applications. Applying GNNs to domains where elements and relationships exist has proven to be very effective. These applications include but are not limited to traffic flow prediction, point cloud classification, and text classification (Cui et al., 2019; Yao et al., 2019; Shi and Rajkumar, 2020). These systems also pose the advantage of implicitly stating the relationships between the different elements, making them potentially more useful for explainability purposes (Pope et al., 2019; Vilone and Longo, 2020, 2021a,b). The use of graphs for EEG-based applications remains underexplored. Zhong et al. (2020) used a Graph Convolutional Network model for EEG-based emotion recognition obtaining results comparable to other non-graph-based methods. Wang et al. (2020) used a graph to map brain functional connectivity for Participant Classification without making use of message-passing Graph Neural Networks. Their work consists of creating brain connectivity networks using multi-channel settings, node centrality, and global network metrics to compute features for the different nodes composing the graph. They prove that generated networks have significant inter-individual distinctiveness, making them suitable for biometric applications.

As these studies suggest, it is possible to obtain meaningful distinguishable patterns of behavior, as gathered by EEG signals, of a specific user, for biometric applications, with current approaches demonstrating error rates close to 0% (Wilaiprasitporn et al., 2019; Seha and Hatzinakos, 2021). However, these approaches rely on datasets composed of several minutes of recordings for each participant, sampling significant EEG windows or using a high number of training samples. Neglecting that sampling data for each participant is an expensive process that could burden real-world scenarios. Carrión-Ojeda et al. (2019) performed a similar study to find out the smallest EEG window size for a biometric system to perform subject identification. The researchers used a fixed amount of training samples per participant, leaving open the question of the minimum total time in seconds required for a PI system to be trained. Similarly, Seha and Hatzinakos (2021) developed a PI model with session-invariant features. Researchers emphasize using smaller windows than previous approaches while obtaining better results. However, they also used most of their available data for training the model, still not focusing on the total training dataset size but solely on the size of the epochs. These reasons motivated the focus on this work to provide hindsight on the minimal amount of affective-based EEG data required from each subject to train task-dependent biometric models. This minimal amount is a combination of the number of samples used for training and the window size for each sample. An exhaustive experiment have been designed, and described in the next section.

This research work is devoted to understanding the minimal EEG data required to identify unique discriminative person-specific features. In order to tackle this goal, the experiment includes a combination of different training data sizes, which are defined based on the size of the sampled EEG windows and the number of training samples available per participant. In detail, these experiments include the training of models by employing several training data sizes along with different machine learning techniques, including a Logistic Regressor (LR), Multi-Layer Perceptron (MLP), 1-D Convolutional Neural Network (CNN), and a Graph Convolutional Network (Graphconv) and a set of feature extraction methods, namely Power Spectral Density (PSD) and Wavelet Energy (WE). Figure 1 depicts the experimental pipeline for a given configuration. Each experimental configuration is used 10 times with different, randomly sampled data to obtain results distributions. Each pipeline aims to train and evaluate a classification model that can perform Person Identification given a sub-sampled window of EEG recordings. Training a discriminative model with data for a pool of p participants that automatically learns to differentiate input features depending on the participant from whom the features have been sub-sampled. More formally, the goal is to approximate a parametric function f_σ that maps an input feature matrix X to a vector of predicted probabilities p with one entry per known participant. p_i represents the probability of the input features belonging to the subject with ID i. All together we can represent it as fσ(X∈ℝn×d)→ps∈[0,1]p where n indicates the number of EEG channels used, d depends on the feature extraction method and denotes the number of obtained features per channel and p is the total number of participants. Models are trained and validated on a randomly sub-sampled portion of the complete dataset, where there are tr_s training samples and 100 validation samples for each user. The evaluation process for a trained model consists of making predictions on the remaining unseen data and measuring classification accuracy. In detail, classification is correct when the highest value for a predicted probability vector p_s matches the ID of the subject the recordings were sampled from.

We hypothesize that by running a set of experimental configurations and evaluating the results, it is possible to provide an answer to the research question stated in Section 1. The overall experiment, consisting of the different configuration settings, is performed to test the following hypothesis:

In the context of training person identification models, the minimum required amount of data is defined based on the total number of train EEG windows used (tr_s) and the size for each window in seconds (w/128). The minimal amount of data and the best learning approach and feature extraction combination can be determined by comparing the result distributions and picking the experimental configuration able to obtain +95% and +99% accuracy on unseen data with the least amount of total training data. Moreover, performing a Mann–Whitney U-test over the results distributions where training data time is the same but the window size differs would answer whether it is better to have fewer oversized windows or many smaller ones.

This research work's main objective is to provide insight into the minimal amount of data required for training an accurate Person Identification model. Each section provides results for an experimental phase. The first phase aims at testing the hypotheses that consider an explicit feature extraction block. This first phase also aims to test the hypothesis of having more minor windows over fewer bigger ones. The second phase tests the hypothesis with no explicit feature extraction block. Results are presented in these two first sections and interpreted in the subsequent discussion section.

Biometric systems using EEG data have seen a rise in popularity with recent advances in ML and DL, along with a deeper understanding of feature extraction methods. These techniques have allowed researchers to obtain low error rates in a wide variety of settings including using recordings sampled from different sessions and having a larger pool of unique subjects. However, there exists little work exploring how much data is needed to train accurate person identification models. This work focuses on answering this question. For this purpose a set of experiments was run comparing different train data sizes, learning approaches and feature extraction methods. The predictive performance of the trained models is then measured based on their accuracy for the unseen, test dataset. This work demonstrates that, in the context of affective EEG-based person identification, having just 1 s of data per participant is enough for training a model to achieve +95% test accuracy. Similarly, 3 s of data suffices to train a model with +99% test accuracy, if a suitable feature extraction method is chosen. The use of raw data is also explored comparing three simple models against a Graph Neural Network which is able to achieve +95% test accuracy using 32 s and +99% test accuracy using 128 s. EEG-based biometrics poses several advantages over traditional biometric systems, the main one it being secure, however, it also poses several disadvantages. Sampling EEG data is not as straightforward as scanning an iris or a fingerprint, several electrodes have to be attached to each subject and their signals must be recorded for a certain amount of time. Data collection is an expensive process. If real-world EEG-based biometric applications were to be implemented, data for every subject would have to be gathered and this data collection process should be as efficient as possible. Current research work focuses mainly on the method but fails to address this fact. These results pave the way for EEG biometrics used in real world applications, the objective being providing insight on how the data quantity changes the performance of the predictive models. Our findings suggest that in order to train an affective EEG-based biometric model, sampling data for a few seconds for every subject would be enough, as opposed to other methods found in literature which use several minutes from each subject to train their respective models. Each subject would not need to sit in a room with electrodes attached to its head for more than a few seconds. As a future direction, the number of electrodes attached to every subject to measure brain signals could also be studied, making the sampling process for this kind of application more straightforward and hence more efficient. Furthermore, task-independent methods considering a wider variety of tasks and session invariant features considering the non-stationary nature of EEG signals on multi-session datasets could also be subject to exploration.

Publicly available datasets were analyzed in this study. This data can be found here: https://www.eecs.qmul.ac.uk/mmv/datasets/deap/.

The main research work was developed by CG-T. LL and BB provided great insight and feedback throughout the project as well as revising the different article drafts that were developed, providing suggestions for potential improvements. All authors contributed to the article and approved the submitted version.

This publication has emanated from research supported in part by a grant from SFI Centre for Research Training in Machine Learning at Technological University Dublin under Grant number 18/CRT/6183. For the purpose of OpenAccess, the author has applied a CC BY public copyright licence to any Author Accepted Manuscript version arising from this submission. We thank Giuliano Anselmi from IBM for granting us access to computing resources and helping us configure the IBM powerstations our models were trained on.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.