Biometric recognition using deep learning

Abstract

The evolution of biometrics has shifted the paradigm of human authentication from classical token and knowledge-based systems to physiological and behavioral trait based systems. Biometric systems have emerged as a next-generation technological solution to strengthen the social and national security. In recent times, biometrics based on brain signal, finger vein pattern and other physiological features have emerged as a futuristic technology which are more fraud resistant compared to the conventional biometrics such as fingerprints. On the other hand, in past few years there has been an increasing interest in application of deep learning techniques in numerous fields including biometrics as it significantly reduces the need for feature engineering. In this dissertation physiological biometric trait such as brain signals and finger vein patterns along with facial feature extraction have been exploited using conventional and deep learning techniques for biometric recognition. Brain signals which are acquired using electroencephalographic (EEG) systems have been long supposed to contain features characteristic of each individual, yet a substantial interest for exploiting them as a potential biometrics for people recognition has only recently grown. The biggest advantages of EEG-based biometrics lie in its universality and security, while its major concerns are related to the acquisition protocol that can be inconvenient and time consuming. In this dissertation we investigate the permanence issue of EEG signals, elicited by visual stimuli, for biometric recognition purposes. Specifically, we evaluate the discriminative capabilities of generic visually-evoked potentials (VEPs) and of visual event-related potentials (ERPs) associated to specific cognitive tasks such as imaginary limbs movement. Furthermore, we analyze the permanence issue of the considered EEG traits by verifying the stability across time of the achievable recognition rates. A deep learning method such as convolutional neural network (CNN), is used for automatic discriminative feature extraction and individual identification. Experimental tests performed on a longitudinal database, comprising of EEG data collected during different sessions, give an evidence of the presence of repeatable discriminative characteristics in the individuals’ EEG activity. Finger vein pattern based human identification has gained a lot of attention in recent years and it is considered to be one of the most promising biometric traits. The performance of the current state-of-the-art techniques, although promising, but are strongly dependent on the quality of the finger vein images, and it may significantly vary for different databases. In this dissertation a convolutional-neural-network (CNN)-based finger vein identification system is proposed, and the performance of the proposed deep neural network is investigated over four publicly available databases. The main purpose of this investigation is to find a standard procedure for finger vein biometric recognition based on deep learning, to be used in stable and highly accurate identification systems, able to cope with different kinds of finger vein images, irrespective of their quality. The results obtained from a set of exhaustive testing show accuracy higher than 95% for all the four considered databases using the proposed CNN-network-based approach which, in comparison with most state-of-the-art approaches, demonstrates the efficiency of the designed CNN network for finger vein identification. In addition to these two physiological traits experimental investigations are performed for facial landmark points estimation using a coarse to fine deep recurrent neural network to accurately extract key landmark point based facial features in order to boost the face recognition accuracy. Experimental results shows that the proposed landmark point estimation techniques is able to achieve a significantly better performance than all the state-of-the-art techniques for the challenging test subset of 300-W competition.