الفهرس 
Biometrics RecognitionOnly 14 pages are availabe for public view
 |  | from | 112 |  |  |
| | | from | 112 | | |
Abstract
The electroencephalogram (EEG) signal is a human biometric that has been extensively used in medical applications. In recent years, the research community has become interested in using EEG signals for identification and security authentication due to their dynamic nature and individuality. Unfortunately, the use of EEG methodology in person identification and authentication is hard because of the complexity of EEG-signal traits and analysis in practice. Moreover, conventional classification models utilized for EEG-person identification heavily rely on signal pre-processing and hand-designed features. To cope with these issues, deep neural networks have been recently employed for EEG classification tasks and show significant result. However, the successes of deep learning algorithms in motor imagery (MI) brain-computer interfaces (BCIs) to increase classification performance is still limited.
In the proposed work, we investigated that the deep learning techniques are more effective to identify persons using EEG signals. The BCI2000 dataset which contains 109 people is used to train three discriminative (CNN, LSTM, and GRU) and two hybrid (1D-convolutional LSTM and 1D-convolutional GRU) deep learning models for person identification.
In the first experiment of the three discriminative models, the obtained results of the accuracy for LSTM model are better than CNN and GRU models, with an accuracy rate of 97.83 %, 96.17 %, and 96.53 %, respectively. In the second experiment of the two hybrid models, it is carried out on three groups of datasets, the results of first group that uses all the 14 experimental EEG recordings data for each subject showed that the accuracy of 1D- convolutional LSTM model is higher
than 1D– convolutional GRU model which achieved accuracy rate of 98.13% and 97.24%, respectively. In the second group, we used 8 experimental recordings data of the motor actions to decrease the dimensionality of the data, the results show that 1D- convolutional LSTM model achieve higher accuracy of 99.2 % than the 1D– convolutional GRU model which achieved accuracy of 97.2 %. In the third group, we used the data of 8 experimental recordings of imaginary tasks to apply our system to identify the disabled people, the results show that the accuracy of 1D- convolutional LSTM model is higher than 1D– convolutional GRU model with achieved accuracy rate of 98.67 %, and 97.19 %, respectively.
In conclusion, the combination of CNNs and LSTMs models in the proposed 1D-convolutional LSTM identification system can greatly improves the accuracy of user identification by using the spatiotemporal features of the EEG signals. Moreover, LSTM is more effective than GRU at extracting latent unique information from large-scale training datasets.
Keywords:
Electroencephalogram (EEG), Brain-Computer Interfaces (BCIs), Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Units (GRU)