Karina Knudsen defended her Master Thesis on Improving Fingerprint Presentation Attack Detection using Deep Learning and Computer Vision
The purpose of this thesis is to explore different options that can help improve the performance of fingerprint presentation attack detection (FPAD) methods. Two specific approaches are chosen due to their novelty and explained in depth from a theoretical perspective and an attempt at an implementation is made. First of all, a regular feed forward neural network (FFNN) is trained to learn how to classify images of fingerprints to either be real or fake. This is done based on certain features that are extracted from known computer vision methods. Another strategy, based on the latest research in deep learning, is improving FPAD through expanding on the training data using a deep convolutional generative adversarial network (DCGAN). Theoretically, the DCGAN should be able to generate artificial images of both real and fake fingerprints, which in turn can be used to train a convolutional neural network (CNN) for classification. The motive behind this method is first of all that increasing the amount of training data should improve the performance of the CNN and thus help make the most of the data available. Second of all, it also addresses the privacy concerns related to storing a large database of fingerprints as fingerprint images generated by a DCGAN does not belong to a person who can be compromised. However, despite being founded on a strong theoretical basis, the implementations of these two models turns out not to yield results as expected. Particularly the DCGAN encounters issues that are known from research to be challenging in order to make the DCGAN converge. Therefore, different techniques on how to mitigate these problems are proposed for further work instead.