CNN-based continuous authentication for digital therapeutics using variational autoencoder

Digital therapeutics (DTx) can be used in conjunction with wearable devices to continuously collect, transmit, and analyze patients’ physiological data, achieving personalized precision medicine. In practice, security and privacy problems may arise when physiological data are improperly protected. For instance, any inaccuracies exist in the collected data (due to malicious attacks or mistakes) may affect the effectiveness of treatment. Additionally, physiological data are characterized by a small amount, resulting in decreased authentication accuracy of deep learning models. Inspired by these observations, this paper presents a continuous authentication scheme for digital therapeutics using variational autoencoder and convolutional neural network (VAECNN). Typically, in the training phase, to improve the stability of CNN model, the collected data are used for data augmentation using VAE. To train a one-class classifier, optimal features are selected by conducting principal component analysis (PCA). During the continuous authentication phase, the trained CNN is utilized for feature extraction. Subsequently, the trained one-class classifier is employed to authenticate the user as a legitimate user or an impostor. To assess the performance of VAECNN, we conduct extensive experiments to estimate its performance, the performance of VAE augmentation and the designed CNN, and compare it with several augmentation schemes and representative authentication schemes. Experimental results indicate that our VAECNN achieves the best performance with elliptic envelope (EE) classifier, i.e., it achieves the lowest equal error rate (EER) of 1.06% and the highest accuracy of 98.96%.