Multi-Stream Conformer-Based User Identification System Using 2D CQT Spectrogram Tailored to Multiple Biosignals

Recent studies actively researching user identification utilizing physiological information to verify identity. However, there is a problem with biometric information exposed to the outside of the body being susceptible to forgery and falsification. Therefore, studies on user identification systems based on biometric signals generated as electrical signals inside the body are necessary to strengthen security with high security and continuous authentication. Among the time-frequency (TF) features applied to existing nonlinear biosignals, short-time fourier transform (STFT) cannot simultaneously improve time-frequency resolution, so there is a limit to improving user identification accuracy. Therefore, this paper introduces a customized constant q transform (CQT) for time-frequency resolution adjustment applied to multiple biosignals. Additionally, a multi-stream conformer combining a convolutional neural network (CNN) and transformer is employed to improve user identification accuracy. Test results confirmed that when using CQT features and a multi-stream conformer for electrocardiogram (ECG) and electroencephalogram (EEG) signals, the accuracy of user identification using both ECG and EEG signals was 97.6%, an improvement of 0.8% or more compared to the accuracy of ECG and EEG single biosignals and CNN-based user identification.