Model Construction and Real-Time Analysis of Personal EEG Identification Based on Machine Learning and Convolutional Neural Network

• Main Authors: Chi-Yuan Hsia, 夏啟元
• Other Authors: 蔡孟勳
• Format: Others
• Language: en_US
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/cgi-bin/gs32/gsweb.cgi/login?o=dnclcdr&s=id=%22107NCHU5396049%22.&searchmode=basic

碩士 === 國立中興大學 === 資訊管理學系所 === 107 === Under the development trend of artificial intelligence, biometrics has become a popular technology, which could be applied to various situations, such as finance, public institutions, and customs. Electroencephalography (EEG), a method for research on biometrics, collects electromagnetic waves on specific positions on the scalp and reflects individual brain activity. Much research proved that α band in EEG could distinguish individual differences, and the significance was proven in clinical neurophysiology. In EEG biometrics, complicated electrode channels were used in most research to cover the entire head for collecting brainwave records. Such equipment could not satisfy the requirement for collectability in the application of biometrics. This study mainly develops the verification model with brainwave through Convolutional Neural Network (CNN). A handy EEG collects the static brainwave of participants for 2 minutes. With the Butterworth Low Pass Filter (BLPF) and Short-time Fourier Transform (STFT), brainwave features are selected from the source brainwave signals, and the verification evaluation model is developed with the comparison between several machine learning classifiers and the deep learning CNN model. Two authentication models of individual specific and general models are proposed in this study and Synthetic Minority Oversampling Technique (SMOTE) is used for solving the imbalance problem between personal data and general data so that the research results show favorable effects in various model evaluation indicators. In individual specific model, the selection of brainwave features at 2 second reveals the accuracy 96.80%. In general model, it is necessary to select brainwave for 20 seconds, which is longer than it in individual specific model, but the accuracy is up to 98.58%. The two models show the advantages and disadvantage, but could be chosen the suitable one for verification systems in distinct application.