Study on fingerprint authentication systems using convolutional neural networks

This paper presents a study on the efficiency of using convolutional neural networks for biometric security systems. Fingerprint identification in images is mainly treated and a VisualGeometryGroup-16 architecture integrated in a Siamese network is used. The Siamese network uses two metrics to determine the similarity between two input images, each metric having a certain experimentally determined threshold. The datasets for training, validation and testing were taken from three sources to increase diversity and contain synthetic changes with different levels of alteration. They are processed in such a manner that they can be received as input for the network. This study aims to establish whether using neural networks is reliable for a biometric security system.

I. Introduction

Physiological and behavioral characteristics have always been used to identify an individual. When we refer to this identification approach, the focus is on biometrics. The development of biometrics happened at a fast pace, starting from technologies based on fingerprint, to those based on iris recognition, facial recognition or even biological rhythm or behavioral characteristics such as specific features of a person’s gait. Because we are in the age of speed and technological progress, biometric authentication has become increasingly common, including in access systems in institutions or in public or private spaces, but also in trade. If biometric technologies can improve our lives, it is no less true that they collect data that falls into the wider sphere of personal data and has as its main purpose the security of the individual.