Detection of Iris Presentation Attacks Using Hybridization of Discrete Cosine Transform and Haar Transform With Machine Learning Classifiers and Ensembles

Iris biometric identification allows for contactless authentication, which helps to avoid the transmission of diseases like COVID-19. Biometric systems become unstable and hazardous due to spoofing attacks involving contact lenses, replayed video, cadaver iris, synthetic Iris, and printed iris. This work demonstrates the iris presentation attacks detection (Iris- PAD) approach that uses fragmental coefficients of transform iris images as features obtained using Discrete Cosine Transform (DCT), Haar Transform, and hybrid Transform. In experimental validations of the proposed method, three main types of feature creation are investigated. The extracted features are utilized for training seven different machine learning classifiers alias Support Vector Machine (SVM), Naive Bayes (NB), Random Forest (RF), and decision tree(J48) with ensembles of SVM+RF+NB, SVM+RF+RT, and RF+SVM+MLP (multi-layer perceptron) for proposed iris liveness detection. The proposed iris liveness detection variants are evaluated using various statistical measures: accuracy, Attack Presentation Classification Error Rate (APCER), Normal Presentation Classification Error Rate (NPCER), Average Classification Error Rate (ACER). Six standard datasets are used in the investigations. Total nine iris spoofing attacks are getting identified in the proposed method. Among all investigated variations of proposed iris-PAD methods, the 4
$\times4$
of fragmental coefficients of a Hybrid transformed iris image with RF algorithm have shown superior iris liveness detection with 99.95% accuracy. The proposed hybridization of transform for features extraction has demonstrated the ability to identify all nine types of iris spoofing attacks and proved it robust. The proposed method offers exceptional performances against the Synthetic iris spoofing images by using a random forest classifier. Machine learning has massive potential in a similar domain and could be explored further based on the research requirements.