Mobile QR Code

1. (Department of Software Kwangwoon University/ Seoul, Korea {hyunbinkong, rhkd865}@gmail.com)
2. (Department of Computer Engineering, Kwangwoon University / Seoul, Korea {kjoohyu, ysin0414}@gmail.com, parkcheolsoo@kw.ac.kr )

EEG, Authentication, Bayesian optimization, Deep learning

## 1. Introduction

Personal authentication through biometrics with each person’s unique biological characteristics is safer from theft, loss, or forgery than physical keys or passwords [1]. Authentication using electroencephalogram (EEG) signals has been studied because of its guarantee of strong security. Unlike EEG, some other biometrics, such as fingerprint and face recognition methods, are easy to steal [2,3]. An EEG records the brain activities on the scalp, which are the averaged electrical responses of neurons in the brain [4]. EEG signals have unique patterns corresponding to each person, making it suitable as an authentication system. On the other hand, EEG signals can be affected by caffeine [5], health conditions, or movement artifacts, such as eye-blinking, or biting [6]. These impediments reduce EEG power in specific frequency bands. In addition, EEG signals are nonstationary; hence, they can vary slightly every day [7]. These can hinder the implementation of an EEG-based authentication system. A convolutional neural network (CNN) model was designed as an authentication task and trained day-to-day to overcome these problems. Using 1-D CNN, the time-series EEG features for each person could be extracted, and the maximum pooling can control the number of parameters to prevent excessive parameter concentration on the dense net [8]. The Bayesian optimization method was applied to search for the optimal structure of the CNN model [9]. Although manual searches, grid searches, and random searches are conventionally utilized for CNN model design, Bayesian optimization helps find the hyperparameters efficiently, considering the experimental results.

## 2. Background

This section describes the EEG and Bayesian optimization.

### 2.1 EEG

In neurophysiology, EEG signals are used as indicators to monitor the changes in brain activity, both spatially and temporally [10]. Its response to brain neural activities shows certain patterns varying with the states of consciousness and mental activities [11]. Therefore, EEG signals indicate the level of brain activity as an objective indicator to determine if the brain's activity increases or decreases. Generally, EEG signals are decomposed into delta, theta, alpha, beta, and gamma waves, depending on the frequency band [12]. A delta wave is a brain wave in the 0.5 to 4Hz range, which is dominant in deep sleep with a fully relaxed brain function. Theta waves refer to EEG signals in the 4 to 8 Hz band, generally referring to the mid-state of body and consciousness, sleepiness, and waking. The alpha wave, in the 8 to 13 Hz band, is the basic wave that reflects the stability of the brain in neurophysiology with EEG signals. This wave has been used traditionally to determine the functional state of the left and right hemispheres of the brain for human behavior because it is less affected by hybrid waves. Beta waves are EEG signals in the 13 to 30 Hz band that appear in awakening, active, and stressed states and are also affected by auditory, tactile, and emotional stimuli. Finally, the gamma wave is a brain wave that occurs relatively frequently in the frontal and parietal lobes under intense anxiety and excitement with external consciousness in the 30-50 Hz band [13].

### 2.2 Bayesian Optimization

The search for the optimal hyperparameters in a deep neural network model is the key to improving its performance. Because of the importance of hyperparameters, many methods have been used to optimize them, such as manual search, grid search, and random search [14]. In Bayesian optimization, an objective function is handled as a random function and places a prior probability distribution. The distribution estimates the change in function by evaluating performance. The distribution is updated post-distribution on the function to collect the next query point. The function is estimated from the query points collected by the Gaussian process. This method can yield the optimal value from the estimated function. The hyperparameters are optimized by repeating this process.

## 3. Materials and Methods

This section describes data acquisition, preprocessing, model, and training.

### 3.1 Data Acquisition

EEG data sets were captured from 10 subjects, who were hungry, had not consumed caffeinated foods or beverages for 12 hours and in a relaxed state with their eyes closed for 10 minutes, over a five day period. The EEG signals were captured with a sampling rate of 600Hz and a 0.05Hz ~ 60Hz notch filter using a g.USBamp (from Guger Technologies, Graz, Austria). The channels used for brain waves were Fp1 and Fp2 2 channels, as shown in Fig. 1 [15].

### 3.2 Preprocessing

The recorded data were contaminated by motion artifacts and other ambient noises, such as the power line noise. To reduce the noise, the raw data were bandpass filtered using an FIR filter between 4 to 40Hz, as shown in Fig. 2.

### 3.3 Model

A model was designed based on six 1D CNN layers with the LeakyReLU activation function. Max pooling layers were applied to reduce the dimensions of the network output and prevent excessive parameter concentrations on the dense net. A 0.3 dropout and batch normalization layers were also added to prevent overfitting. The model was trained using a binary cross-entropy loss and an Adam optimizer. Hyperparameters and the number of layers of the model were optimized using Bayesian optimization, whose parameters and search spaces are listed in Table 1.

The search step of Bayesian optimization was 16. The numbers of total parameters and trainable parameters were 70,001 and 69,777. Fig. 3 shows t}he finally optimized model.

##### Table 1. Search spaces of the Bayesian optimization process.
 Layer Search space Input shape 600(Hz) x 3(sec) Conv1D 16 ~ 128 MaxPooling1D 2 Conv1D 16 ~ 128 MaxPooling1D 2 Conv1D 16 ~ 128 MaxPooling1D 2 Dropout 0.3 BatchNormalization - Conv1D 16 ~ 128 MaxPooling1D 2 Conv1D 16 ~ 128 MaxPooling1D 2 Conv1D 16 ~ 128 MaxPooling1D 2 Dropout 0.3 BatchNormalization - Flatten - Dense 16 ~ 128 Dense 16 ~ 64 Dense 1 Activation softmax

### 3.4 Training and Testing Process

The validation set for updating the network weights included 10% of the data recorded each day. For the incremental learning each day, t}he training data sets were prepared from the 1$^{\mathrm{st}}$ day, from the 1$^{\mathrm{st}}$ to the 2$^{\mathrm{nd}}$ day, from the 1$^{\mathrm{st}}$ to the 3$^{\mathrm{rd}}$ day, and from the 1$^{\mathrm{st}}$ to the 4$^{\mathrm{th}}$ day. The test data sets were prepared using the recordings from the 2$^{\mathrm{nd}}$ to the 5$^{\mathrm{th}}$ day. The training and testing were conducted in an inter-subject manner.

##### Table 3. Averaged Testing Results.
 Test data Train data Metrics Day1 Day1~ Day2 Day1~ Day3 Day1~ Day4 Val accuracy 95.66% ± 2.75% 96.65% ± 3.08% 96.91% ± 2.20% 96.64% ± 1.97% precision 85.80% ±15.95% 87.89% ±10.86% 96.53% ±5.07% 87.51% ±13.14% recall 65.88% ±29.01% 75.95% ±23.43% 72.87% ±17.62% 80.65% ±16.29% Day2 accuracy 93.52% ±3.69% - - - precision 62.95% ±31.99% - - - recall 45.80% ±30.83% - - - Day3 accuracy 92.50% ±3.29% 93.22% ±3.43% - - precision 54.94% ±28.29% 69.41% ±17.12% - - recall 42.25% ±32.71% 53.41% ±28.32% - - Day4 accuracy 92.60% ±3.37% 93.31% ±3.49% 93.94% ±3.25% - precision 60.45% ±24.34% 69.91% ±17.64% 80.60% ±16.95% - recall 43.01% ±31.54% 55.51% ±28.77% 47.70% ±28.48% - Day5 accuracy 91.34% ±3.18% 92.76% ±3.25% 93.26% ±2.64% 93.23% ±2.73% precision 65.64% ±21.36% 73.80% ±14.49% 79.61% ±8.09% 71.31% ±12.55% recall 36.13% ±25.65% 51.44% ±23.10% 44.95% ±28.21% 57.65% ±28.82%

## 4. Results

The performance was measured using three metrics: accuracy, precision, and recall which were calculated using Eqs. (2)-(4)

##### (2)
$\textit{Accuracy}=\frac{TP+TN}{TP+FP+TN+FN} \\$
##### (3)
$\textit{Precision}=\frac{TP}{TP+FP} \\$
##### (4)
$\textit{Recall}=\frac{TP}{TP+FN}$

where $\textit{TP, TN, FP,}$ and $\textit{FN}$ denote the true positive, true negative, false positive, and false negative, respectively. Ten different models were designed for each subject as a binary classifier: self-versus the others. Table 3 lists the results with the mean and standard deviation. Each number represents the averaged result across all subjects. The 1$^{\mathrm{st}}$ row elaborates the training dataset prepared from the 1$^{\mathrm{st}}$ day, from the 1$^{\mathrm{st}}$ to the 2$^{\mathrm{nd}}$ day, from the 1$^{\mathrm{st}}$ to the 3$^{\mathrm{rd}}$ day, and from the 1$^{\mathrm{st}}$ to the 4$^{\mathrm{th}}$ day. The 1$^{\mathrm{st}}$ column denotes the validation dataset (the same day dataset as the training dataset}) and testing dataset, which includes the data recorded on the next day of the training dataset, such as the 2$^{\mathrm{nd}}$, 3$^{\mathrm{rd}}$, 4$^{\mathrm{th}}$, and 5$^{\mathrm{th}}$ days.

The results of the validation data showed the training performance of the model using the same day data; all showed an accuracy of more than 90% across all 10 subjects. As shown in Table 3, t}he performance decreased when the recording day of the testing dataset was far from the day of the training dataset. Compared to the standard deviation of the accuracy, those of the precision and recall were relatively higher.

To assess the improvement in performance with increasing training dataset size, the variations in the result with additional training data on the following day were calculated, as listed in Table 4. Overall, the incremental learning of the EEG signals day-to-day enhanced the authentication performance significantly with a decreasing standard deviation.

##### Table 4. Variations with an additional training dataset.
 Test data Train data Metrics Day1~ Day2 Day1~ Day3 Day1~ Day4 Day2 accuracy +0.99% +0.26% -0.27% precision +2.07% +8.63% -9.02% recall +10.06% -3.07% +7.78% Day3 accuracy +0.72% - - precision +14.47% - - recall +11.17% - - Day4 accuracy +0.72% +0.63% - precision +9.47% +10.69% - recall +12.50% -7.80% - Day5 accuracy +1.42% +0.50% -0.03% precision +8.15% +5.81% -8.30% recall +15.30% -6.48% +12.69%

## 5. Conclusion

This paper reported a feasible way to incrementally train nonstationary EEG data for personal authentication using 1D-CNN and Bayesian optimization. The proposed neural network model yielded an averaged accuracy, precision, and recall of 93.23%, 71.31%, and 57.65%, respectively. The different conditions of the subjects every day and EEG deformation could affect the performance. For future work, controlling more variables and additional recordings will be needed to improve the result.

### ACKNOWLEDGMENTS

This research was supported by the MIST (Ministry of Science and ICT), under the National Program for Excellence in SW (2017-0-00096), supervised by the IITP(Institute for Information & communications Technology Promotion)

### REFERENCES

1
Jain A. K., Ross A., Pankanti S., Jun. 2006, Biometrics: A Tool for Information Security, IEEE Transactions on Information Forensics and Security, Vol. 1, No. 2, pp. 125-143
2
Marcel S., Millan J. R., Apr. 2007, Person Authentication Using Brainwaves (EEG) and Maximum A Posteriori Model Adaptation, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 29, No. 4, pp. 743-752
3
Ashby C., Bhatia A., Tenore F., Vogelstein J., 2011, Low-cost electroencephalogram (EEG) based authentication, in 2011 5th International IEEE/ EMBS Conference on Neural Engineering
4
Sanei S., Chambers J., 2007, EEG Signal Processing
5
Dimpfel W., Schober F., Spüler M., Mar. 1993, The influence of caffeine on human EEG under resting condition and during mental loads, The Clinical Investigator, Vol. 71, No. 3
6
Abo-Zahhad M., Ahmed S. M., Abbas S. N., Oct. 2016, A new multi-level approach to EEG based human authentication using eye blinking, Pattern Recognition Letters, Vol. 82, pp. 216-225
7
Marcel S., Millan J. R., Apr. 2007, Person Authentication Using Brainwaves (EEG) and Maximum A Posteriori Model Adaptation, IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 29, No. 4, pp. 743-752
8
Wang M., El-Fiqi H., Hu J., Abbass H. A., Dec. 2019, Convolutional Neural Networks Using Dynamic Functional Connectivity for EEG-Based Person Identification in Diverse Human States, IEEE Transactions on Information Forensics and Security, Vol. 14, No. 12, pp. 3259-3272
9
Močkus J., 1975, On Bayesian methods for seeking the extremum, Optimization techniques IFIP technical conference, pp. 400-404
10
Sun Y., Lo F. P.-W., Lo B., Jul. 2019, EEG-based user identification system using 1D-convolutional long short-term memory neural networks, Expert Systems with Applications, Vol. 125, pp. 259-267
11
Zeynali M., Seyedarabi H., Aug. 2019, EEG-based single-channel authentication systems with optimum electrode placement for different mental activities, Biomedical Journal, Vol. 42, No. 4, pp. 261-267
12
Wang M., El-Fiqi H., Hu J., Abbass H. A., Dec. 2019, Convolutional Neural Networks Using Dynamic Functional Connectivity for EEG-Based Person Identification in Diverse Human States, IEEE Transactions on Information Forensics and Security, Vol. 14, No. 12, pp. 3259-3272
13
Wang J., Barstein J., Ethridge L. E., Mosconi M. W., Takarae Y., Sweeney J. A., Sep. 2013, Resting state EEG abnormalities in autism spectrum disorders, Journal of Neurodevelopmental Disorders, Vol. 5, No. 1
14
Bergstra J., Bardenet R., Bengio Y., Kgl B., 2011, Algorithms for Hyper-Parameter Optimization, Neural Information and Processing Systems
15
Jurcak V., Tsuzuki D., Dan I., Feb. 2007, 10/20, 10/10, and 10/5 systems revisited: Their validity as relative head-surface-based positioning systems, NeuroImage, Vol. 34, No. 4, pp. 1600-1611

## Author

##### Hyeonbin Lee

Hyeonbin Lee is in the BSc Program at the Department of Software at Kwangwoon University, Seoul, Republic of Korea. His research interests include security, machine learning algorithms, deep learning, and reinforcement learning.

##### Gwangho Kim

Gwangho Kim is in the BSc Program at the Department of Software at Kwangwoon University, Seoul, Republic of Korea. His research interests include machine learning algorithms, deep learning, specifically, AutoML, and explainable AI

##### Juhyeong Kim

Juhyeong Kim is in the MSc Program at the Bio Computing & Machine Learning Laboratory (BCML) in the Department of Computer Engineering at Kwangwoon University, Seoul, Republic of Korea. He received a BSc from the Department of Computer Engineering, Kwangwoon University, Seoul, Republic of Korea, in 2019. His research interests include machine learning algorithms, specifically, deep learning, reinforcement learning, and automated machine learning (AutoML)

##### Youngshin Kang

Youngshin Kang graduated from the computer engineering department at Far East University in Chungbuk, South Korea. Her research interests include machine learning, deep learning optimization, biological signal, and signal processing. She is particularly interested in brain engineering.

##### Cheolsoo Park

Cheolsoo Park is an associate professor in the Computer Engineering Department at Kwangwoon University, Seoul, South Korea. He received a B.Eng. in Electrical Engineering from Sogang University, Seoul, and an MSc from the Biomedical Engineering Department at Seoul National University, South Korea. In 2012, he received his Ph.D. in Adaptive Nonlinear Signal Processing from Imperial College London, London, UK. He worked as a postdoctoral researcher in the Bioengineering Department at the University of California, San Diego, USA. His research interests are mainly in machine learning and adaptive and statistical signal processing, with applications to brain-computer interfaces, computational neuroscience, and wearable technology.