Akademik Veri Yönetim Sistemi
TRAITEMENT DU SIGNAL, cilt.40, sa.1, ss.65-79, 2023 (SCI-Expanded)
Owing to the reliability of biometric data, person identification systems are developed using many different biometric data. However, these systems can be easily fooled with prosthetic face masks, contact lenses and fingerprint tapes. EEG signal is considered to be the most difficult biometric data to copy. The main reason why EEG-based identification systems have not become widespread is that their accuracy performance is not stable. In this study, an EEG based identification system DM-EEGID with improved accuracy performance is proposed. In this approach, first of all, the channels that are meaningless and reduce the accuracy performance from the high number of channels used as input data should be filtered out. Therefore, a Random Forest based binary feature selection method is recommended. With this algorithm, it has been determined that the optimum number of channels for the highest percentage of accuracy in the 64-channel data set is 48-channel. Then, for the determination of the most distinctive frequency subcomponent, the delta pattern was determined to be the most appropriate frequency component by inter-section correlation coefficient analysis. Finally, the proposed approach was tested with hybrid Attention-based LSTM-MLP supported by optimum parameters with both eyes closed and eyes open resting state EEG recording. The proposed model reached 99.96% and 99.70% accuracy percentages for eyes-closed and eyes-open datasets, respectively. These results show that this proposed approach has the potential to be applied in closed systems where the number of people is limited.