Deep Transfer Learning Model for EEG Biometric Decoding

Abstract

In automated systems, biometric systems can be used for efficient and unique identification and authentication of individuals without requiring users to carry or remember any physical tokens or passwords. Biometric systems are a rapidly developing and promising technology domain. in contrasting with conventional methods like password IDs. Biometrics refer to biological measures or physical traits that can be employed to identify and authenticate individuals. The motivation to employ brain activity as a biometric identifier in automatic identification systems has increased substantially in recent years. with a specific focus on data obtained through electroencephalography (EEG). Numerous investigations have revealed the existence of discriminative characteristics in brain signals captured during different types of cognitive tasks. However, because of their high dimensional and nonstationary properties, EEG signals are inherently complex, which means that both feature extraction and classification methods must take this into consideration. In this study, a hybridization method that combined a classical classifier with a pre-trained convolutional neural network (CNN) and the short-time Fourier transform (STFT) spectrum was employed. For tasks such as subject identification and lock and unlock classification, we employed a hybrid model in mobile biometric authentication to decode two-class motor imagery (MI) signals. This was accomplished by building nine distinct hybrid models using nine potential classifiers, primarily classification algorithms, from which the best one was finally selected. The experimental portion of this study involved, in practice, six experiments. For biometric authentication tasks, the first experiment tries to create a hybrid model. In order to accomplish this, nine hybrid models were constructed using nine potential classifiers, which are largely classification methods. Comparing the RF-VGG19 model to other models, it is evident that the former performed better. As a result, it was chosen as the method for mobile biometric authentication. The performance RF-VGG19 model is validated using the second experiment. The third experiment attempts for verifying the RF-VGG19 model's performance. The fourth experiment performs the lock and unlock classification process with an average accuracy of 91.0% using the RF-VGG19 model. The fifth experiment was performed to verify the accuracy and effectiveness of the RF-VGG19 model in performing the lock and unlock task. The mean accuracy achieved was 94.40%. Validating the RF-VGG19 model for the lock and unlock task using a different dataset (unseen data) was the goal of the sixth experiment, which achieved an accuracy of 92.8%. This indicates the hybrid model assesses the left and right hands' ability to decode the MI signal. Consequently, The RF-VGG19 model can aid the BCI-MI community by simplifying the implementation of the mobile biometric authentication requirement, specifically in subject identification and lock and unlock classification.