Alexandru-Cosmin Vasile defended his Master Thesis on Evaluation of Deep Learning Methods for Fingerprint Presentation Attack Detection
The usage of biometric systems for automating identification and verification of individuals represents an increasing trend. It is critical that the security of those systems is optimal in order to prevent misuse. Presentation Attack Detection (PAD) methods need to be employed for an increased security on various types of biometric data such as fingerprints. This thesis explores the feasibility of using deep learning techniques as PAD methods, working on fingerprint Laser Speckle Contrast Imaging (LSCI) data. The focus is on utilising Long Short-Term Memory (LSTM) networks in order to effectively process the temporal information that is present in the data. Fourteen deep neural network models are implemented and evaluated on a private data set comprising LSCI fingerprint data. The data set comprises 2302 samples, with 1146 Presentation Attack (PA) samples and 1156 bona fide samples. Five of the trained and evaluated models achieve a classification accuracy above 97%. The best performing model achieves a classification accuracy of 97.86%, with an Attack presentation classification error rate (APCER) of 1.92% and Bona fide presentation classification error rate (BPCER) of 2.35%. The study concludes that it is imperative to use a convolutional base together with a LSTM block in order to achieve high classification accuracy. Out of the 36 Presentation Attack Instrument (PAI) species present in the data set, the overlay type represents the biggest challenge in classification for all evaluated models.