Investigate the Effect of Sample Synthesis on Deep Networks based Iris Liveness Detection
Hui Zhang, Jing Liu, Dexin Zhang, Bangyu Li · 2024
In addition to using synthetic samples to deceive and attack biometric systems, they can also be used to enhance algorithmic robustness. This study investigates the impact of synthetic samples on cosmetic contact lens detection, a typical iris classification task that presents a bottleneck problem. By incorporating varying amounts of synthetic images into the training dataset and training multiple classifiers, we analyze the effect of synthetic samples. For iris liveness detection, we propose a novel SACvT network that incorporates spatial local attention to improve the CvT. The SACvT network comprises the CvT as its backbone and a spatial local attention network that generates an attention feature map from the last Token map of the CvT. A Rank loss is designed to train SACvT without requiring attention annotations. We choose two light-weighted networks as comparison methods, including a modified LightCNN and the CvT. Extensive experiments show that adding synthetic samples based on StyleGANv2 for training can improve classification accuracy. While GAN-based data augmentation cannot replace image processing-based data augmentation, it serves as a complementary approach. The proposed SACvT exhibits promising results in liveness detection, particularly on challenging images such as those with imprecise iris localization or heavy occlusion.