Yap Hui Yen, Choo Yun-Huoy, Mohd Yusoh Zeratul Izzah, Khoh Wee How
Faculty of Information Science and Technology, Multimedia University (MMU), Melaka, Malaysia.
Faculty of Information and Communication Technology, Universiti Teknikal Malaysia Melaka (UTeM), Melaka, Malaysia.
Brain Inform. 2023 Aug 3;10(1):19. doi: 10.1186/s40708-023-00198-4.
Electroencephalogram(EEG)-based authentication has received increasing attention from researchers as they believe it could serve as an alternative to more conventional personal authentication methods. Unfortunately, EEG signals are non-stationary and could be easily contaminated by noise and artifacts. Therefore, further processing of data analysis is needed to retrieve useful information. Various machine learning approaches have been proposed and implemented in the EEG-based domain, with deep learning being the most current trend. However, retaining the performance of a deep learning model requires substantial computational effort and a vast amount of data, especially when the models go deeper to generate consistent results. Deep learning models trained with small data sets from scratch may experience an overfitting issue. Transfer learning becomes an alternative solution. It is a technique to recognize and apply the knowledge and skills learned from the previous tasks to a new domain with limited training data. This study attempts to explore the applicability of transferring various pre-trained models' knowledge to the EEG-based authentication domain. A self-collected database that consists of 30 subjects was utilized in the analysis. The database enrolment is divided into two sessions, with each session producing two sets of EEG recording data. The frequency spectrums of the preprocessed EEG signals are extracted and fed into the pre-trained models as the input data. Three experimental tests are carried out and the best performance is reported with accuracy in the range of 99.1-99.9%. The acquired results demonstrate the efficiency of transfer learning in authenticating an individual in this domain.
基于脑电图（EEG）的身份认证越来越受到研究人员的关注，因为他们认为它可以作为更传统的个人身份认证方法的替代方案。不幸的是，EEG信号是非平稳的，很容易被噪声和伪迹污染。因此，需要进一步进行数据分析处理以检索有用信息。在基于EEG的领域中已经提出并实施了各种机器学习方法，深度学习是当前最流行的趋势。然而，要保持深度学习模型的性能需要大量的计算工作和大量的数据，特别是当模型变得更深以产生一致的结果时。从零开始用小数据集训练的深度学习模型可能会出现过拟合问题。迁移学习成为一种替代解决方案。它是一种识别并将从先前任务中学到的知识和技能应用于训练数据有限的新领域的技术。本研究试图探索将各种预训练模型的知识迁移到基于EEG的身份认证领域的适用性。分析中使用了一个由30名受试者组成的自收集数据库。数据库注册分为两个阶段，每个阶段产生两组EEG记录数据。提取预处理后的EEG信号的频谱，并将其作为输入数据输入到预训练模型中。进行了三项实验测试，报告的最佳性能准确率在99.1%-99.9%之间。获得的结果证明了迁移学习在该领域进行个人身份认证方面的效率。
