Presentation Attack Detection for Mobile Device-Based Iris Recognition

Apart from ensuring high recognition accuracy, one of the main challenges associated with mobile iris recognition is reliable Presentation Attack Detection (PAD). This paper proposes a method of detecting presentation attacks when the iris image is collected in visible light using mobile devices. We extended the existing database of 909 bonafide iris images acquired with a mobile phone by collecting additional 900 images of irises presented on a color screen. We explore different image channels in both RGB and HSV color spaces, deep learning-based and geometric model-based image segmentation, and use Local Binary Patterns (LBP) along with the selected statistical images features classified by the Support Vector Machine to propose an iris PAD algorithm suitable for mobile iris recognition setups. We found that the red channel in the RGB color space offers the best-quality input samples from the PAD point of view. In subject-disjoint experiments, this method was able to detect 99.78% of screen presentations, and did not reject any live sample.