Iris presentation attack detection based on best-kfeature selection from YOLO inspired RoI

摘要

Obfuscating an iris recognition system through forged iris samples has been a major security threat in iris-based authentication. Therefore, a detection mechanism is essential that may explicitly discriminate between the live iris and forged (attack) patterns. The majority of existing methods analyze the eye image as a whole to find discriminatory features for fake and real iris. However, many attacks do not alter the entire eye image, instead merely the iris region is affected. It infers that the iris embodies the region of interest (RoI) for an exhaustive search towards identifying forged iris patterns. This paper introduces a novel framework that locates RoI using the YOLO approach and performs selective image enhancement to enrich the core textural details. The YOLO approach tightly bounds the iris region without any pattern loss, where the textural analysis through local and global descriptors is expected to be efficacious. Afterward, various handcrafted and CNN based methods are employed to extract the discriminative textural features from the RoI. Later, the best-kfeatures are identified through the Friedman test as the optimal feature set and combined using score-level fusion. Further, the proposed approach is assessed on six different iris databases using predefined intra-dataset, cross-dataset, and combined-dataset validation protocols. The experimental outcomes exhibit that the proposed method results in significant error reduction with the state of the arts.