Unsupervised Autoencoders for Generalized Iris Presentation Attack Detection

Iris Presentation Attacks (PAs) such as print-scan, contact lenses, synthetic iris images, etc. poses threat to current iris recognition algorithms. Many researchers have been working to solve this problem, but their approaches are restricted to only seen distribution of attacks. Only handful of researchers are working on building a generalized algorithm for both seen and unseen attacks. Recently, researchers have proposed use of unsupervised autoencoders for anomaly detection in image classification. Using similar notion, I am working on building an unsupervised Autoencoder for spoof detection. Here autoencoders are trained using only bona fide samples, learning their distribution that can be used to differentiate real iris images from PAs