Detecting Images Outside Training Distribution for Fingerprint Spoof Detection

Abstract: Artificial neural networks are known to run into issues when given samples that deviate from the training distribution, where the network may confidently provide an incorrect answer. Out-of-distribution detection methods aims to provide a solution to this issue, by detecting data that deviates from the distribution used to train the model. This thesis looks at the possibility of using out-of-distribution methods in a more challenging context, where the data looked at is more semantically similar than what is often looked at in the literature. Three out-of-distribution detection methods are tested and evaluated at separating fingerprint images from out-of-distribution images, including spoof molds of fingerprints, generated fingerprints using GANs, and non-finger images. The results on the non-finger dataset are in line with the literature, and all methods give promising results on the non-finger dataset. None of the methods are however able to separate the more challenging spoof molds dataset from the in-distribution images. However, on the generated dataset the methods are able to provide somewhat encouraging results, though the performance is significantly lower than on the semantically different datasets.