GAN-Based Data Augmentation for Improving Biometric Authentication Using CWT Images of Blood Flow Sounds

Biometric identification allows to secure sensitive information. Since existing biometric traits, such as finger-prings, voice, etc. are associated with different limitations, we exemplified the potential of blood flow sounds for biometric authentication in previous work. Therefore, we used measurements from seven different users acquired with a custom-built auscultation device to calculate the spectrograms of these signals for each cardiac cycle using continuous wavelet transform (CWT). The resulting spectral images were then used for training of a convolutional neural network (CNN). In this work, we repeated the same experiment with data from twelve users by adding more data from the original seven users and data from five more users. This lead to an imbalanced dataset, where the amount of available data for the new users was much smaller, e.g., U1 had more than 900 samples per side whereas the new user U9 had less than 100 samples per side. We experienced a lower performance for t he new users, i.e. their sensitivity was 18-21% lower than the overall accuracy. Thus, we examined whether the augmentation of data leads to better results. This analysis was performed using generative adversarial networks (GANs). The newly generated data was then used for training of a CNN with several different settings, revealing the potential of GAN-based data augmentation for increasing the accuracy of biometric authentication using blood flow sounds.