Statistical representation for iris anti-spoofing using wavelet-based feature extraction and selection algorithms

The development of fake iris detection systems, which is one of the most important topics in the biometric field, is growing rapidly. In this paper, discriminative statistical features are used for differing between real and fake iris images. The multilevel 2-D wavelet decomposition is employed to obtain approximation and detail wavelet channels. For feature classification, Euclidean distance and suitable fusion rules are applied. Problems with numerous features require the use of feature selection. Thus, to reduce the computational cost and enhance the system performance, an effective feature selection algorithm is proposed. CASIA-Iris-Syn database, which consists of about 10000 synthesized images, is used. Results show that the variance measure is efficient for detecting deceived attacks with 100% classification accuracy. The kurtosis measure gives 90.4648 %, which is the lowest accuracy obtained. Other feature selection algorithms are applied for a comparison purpose. Results prove the high explanatory capability of the prediction method. The proposed feature selection/classifier ensemble not only achieves dimensionality reduction, but also carefully investigates the dependence between the statistical features, and does not neglect features with complementary information. A poor choice of features may lead to significant deterioration in system performance. The proposed system has the advantage of working with large size database, and thus ensures the generalization ability of the proposed algorithms. Results also show that working with original non-segmented images not only reduces the processing time, but also enhances the classification accuracy, noticeably.