Deep Learning Evaluation Using Receiver Operating Curve (ROC) for Footprint Biometric Authentication

Abstract

Biometric authentication compares two templates of individual biological characteristics, one of which is submitted by the user and the other is retrieved from the system. Among the comparison techniques that have been widely used in biometric systems, is deep learning. When deep learning is used for authentication, there is a conflict regarding the performance evaluation criteria of the system. The criteria that are usually used to evaluate the performance of deep learning models are accuracy and loss. It is obtained from the analysis of training and verification data submitted by an individual. In this case, the validation data comes from users, which is known by the deep learning model. In authentication terms, this measurement is known as the ‘Genuine’ score. This evaluation is still insufficient to conclude the performance of a deep learning model. The ‘Imposter’ score also needs to be considered so that the evaluation of this system can be done comprehensively. The focus of this paper is on the performance evaluation procedure of deep learning models for authentication, considering the ‘Imposter Score’ when constructing the Receiver Operating Characteristic (ROC) curve. Footprints have been used as a biometric feature in this study. The AlexNet architecture was chosen to train the system. 45 ‘Genuine’ scores and 2880 ‘Imposter’ scores were successfully generated from 15 individuals using the proposed procedure. At the end of this study, it was found that when the ‘Imposter’ score is included when evaluating the deep learning model, it can give a better conclusion about the deep learning authentication performance. The training duration of a deep learning model can also be reduced based on discriminative authentication criteria.