Protected template classification for iris biometrics

Although the wide application of biometrics has brought much convenience to the daily lives of people, it has also resulted in several security risks. Currently, many classical attack methods assume that the attacker has mastered all details of the template protection scheme. However, in practical environments, attackers often find it difficult to obtain the complete system information. Therefore, this paper proposes a two-stage classification model that can effectively classify template protection schemes. The model adopts a two-stage classification strategy: the first stage focuses on the overall classification of template protection schemes, whereas the second stage analyses the specific implementation details of each scheme and its related parameter settings in detail. Through this design, template protection schemes can still be effectively evaluated and attacked, even if the attacker does not have a full understanding of the specific details of the system. Further, this study explores the role of different classifiers in classification models in detail to help select the most appropriate classifier for improving the performance of the model. We compare several common classifiers and analyse their performance on different datasets. The experimental results show that the deep learning models (ResNet18 and DenseNet) outperform traditional machine learning models (LR, DT, RF, ADA) on all the datasets, with stable F1 scores greater than 0.90 on several testing sets, whereas the F1 scores of traditional machine learning models are generally low. In particular, on the CASIA-IrisV3-Interval and CASIA-IrisV4-Lamp datasets, DenseNet achieves F1 scores of 0.91-0.98, showing excellent generalisation ability. Further, the experiments show that deep learning models maintain high classification accuracies (F1 scores mostly higher than 0.90) even when the data sources of the template protection schemes differ, whereas the performance of some traditional models (e.g., ADA) fluctuates greatly on different datasets. This result suggests that with sophisticated guards, attackers can leverage the powerful classification capabilities of deep learning models to effectively analyse biometric data, thus providing a solid foundation for subsequent attack steps.