Iris Liveness Detection by Relative Distance Comparisons

Abstract

The focus of this paper is on presentation attack detection for the iris biometrics, which measures the pattern within the colored concentric circle of the subjects' eyes, to authenticate an individual to a generic user verification system. Unlike previous deep learning methods that use single convolutional neural network architectures, this paper develops a framework built upon triplet convolutional networks that takes as input two real iris patches and a fake patch or two fake patches and a genuine patch. The aim is to increase the number of training samples and to generate a representation that separates the real from the fake iris patches. The smaller architecture provides a way to do early stopping based on the liveness of single patches rather than the whole image. The matching is performed by computing the distance with respect to a reference set of real and fake examples. The proposed approach allows for real-time processing using a smaller network and provides equal or better than state-of-the-art performance on three benchmark datasets of photo-based and contact lens presentation attacks.