Abstract:
Biometric authentication using Electroencephalo-gram (EEG) signals is getting more popular because it is hard to fake or steal as compared to things like fingerprints, iris or face scans. Electroencephalogram(EEG) is unique to each person and we cannot just sneak up and generate or grab those brain waves easily. But most of the systems nowadays use these handpicked features from the spectrum and they only test on a few people. So it does not really work well for everyday use. In this paper, we came up with EEGAuthNet, a deep learning based setup that takes raw EEG signals from 64 channels and authenticates without the need to extract features manually.The model combines one-dimensional Convolutional Neural Networks (CNN) to pull out spatial features and then it adds Bidirectional Long Short-Term Memory (Bi-LSTM) layers that go both ways to handle the longer patterns in time across the EEG bits. We tested it on PhysioNet dataset for motor movements and imagery which contains 109 subjects.The accuracy for identifying came out to be 81.47%. Equal Error Rate (ERR) was 2.0%. To check what each part does, we did an ablation study which confirmed that each component contributes meaningfully with combination of CNN and Bi-LSTM which beats only CNN by 3.8Overall this shows EEG could be a good secure way for biometrics if you pair it with this hybrid setup.
Keywords: