From the Perspective of CNN to Adversarial Iris Images

Abstract

IrisCode, the most influential iris recognition algorithm, has been extensively applied in national identity programmes and border controls. Currently over one billion people spanning approximately 170 nationalities have been enrolled in IrisCode. It is vital to thoroughly study this crucial algorithm, especially for issues related to its security. How to generate iris images from its templates is one of such security issues. The existing generation methods can produce images from iris templates and the synthetic images can be used to match other real iris images. However, their quality is very low with obvious artifacts. Considering the feature extraction process of IrisCode as a non-differentiable shallow convolutional neural network and using a differentiable function to approximate its step function, the generation process can be formulated as an unconstrained minimization problem. It can be mathematically proven that the solution of this formulation is the same as that of the convex polyhedral cone method proposed before. By exploiting the unconstrained formulation and the step function directly, a constrained convex minimization formulation with an inputted iris image selected by a search function as an information carrier is derived. The experimental results on the UBIRIS.v1 database and the WVU database demonstrate that the proposed algorithm can generate high quality iris images and significantly outperform the previous methods in terms of visual quality and six image quality metrics. They can also be matched with other real iris images in the two databases.