Researchers progress further on iris biometric liveness detection with multi-class networks
A three-class serial model for presentation attack detection to protect iris biometric systems shows benefits over the two-class approach that was proven effective in international competition.
A paper on ‘Iris Liveness Detection Using a Cascade of Dedicated Deep Learning Networks’ describes the technique that won the LivDet-Iris 2020 competition, and extends it with three and four-class models. The paper was authored by Juan Tapia Farias, Sebastián González Sandoval, and Christoph Busch, and presented this week at the International Joint Conference on Biometrics (IJCB 2022) in Abu Dhabi.
The researchers built a large iris presentation attack dataset, and focussed on detecting bona fide images. This contrasts with existing studies on iris PAD, which tend to address a specific attack vector.
The two-class model achieved BPCER10 (Bona Fide Classification Error Rate) of 0.99 percent, the three-class model returned 0.16 percent, and the four-class model scored 0.83 percent. For BPCER20, the values for the two- and four-class models rose above 3 percent, but the three-class model again returned 0.16 percent.
The research also revealed that input images 224 by 224 pixels are adequate to detect bona fide irises, but PAD results improve with 448 by 448-pixel images.
The report authors used aggressive data augmentation to train modified MobileNetV2 networks, and ran five experiments with networks fine-tuned or trained from scratch.
“When trained from scratch, our suggested networks allow us to complement the results of the LivDet-Iris 2020 competition by using more challenging PAI species,” the researchers conclude. “When using fine-tuning, model performance worsens in proportion to the number of layers from the network that were frozen. Nonetheless, results using fine-tuning are competitive with the literature.”