Enhanced iris presentation attack detection via contraction-expansion CNN

Abstract

• Proposed a novel contraction-expansion CNN architecture for iris presentation attack detection. • Experiments in a generalized setting are performed to replicate real-world scenarios. • Extensive comparisons showcase the excellence of the proposed algorithm. The vulnerability of iris recognition algorithms against presentation attacks including textured contact lens is a serious concern. Several presentation attack detection (PAD) algorithms are developed; however, most of them suffer from low generalization capability. In this research, we propose a two head ‘contraction expansion’ convolutional neural network (CNN) architecture for robust presentation attack detection. The architecture’s input consists of raw image and edge enhanced image to learn the discriminating features for binary classification . The experiments using multiple iris presentation attack databases, including the LivDet-2017 and IIITD contact lens database (CLI), showcase the efficacy of the proposed algorithm. For example, the proposed iris presentation attack detection (IPAD) network yields 11.1% lower average classification error rate than recent state-of-the-art algorithm namely MVANet on the UnMIPA database when the system is trained on an unseen IIITD CLI database.