Master thesis presentation by Jenny Nilsson, MPBME: “Feature-based quality assessment for spoof fingerprint images”

Welcome to a master thesis presentation with the title “Feature-based quality assessment for spoof fingerprint images” by Jenny Nilsson, MPBME.

When? Friday, 20 January at 13.15 pm.

Where? Lunnerummet (room 3311), Hörsalsvägen 11, 3rd floor, Chalmers

Examiner: Fredrik Kahl

 

Abstract

Fingerprint recognition has over the last decade become a natural component in modern identity management systems. As the commercial use of fingerprint recognition systems increases, the benefits from attacking such systems becomes greater. The security of a biometric system is seriously compromised if the system is unable to differentiate between a real and a counterfeit fingerprint. From this security threat a need for methods to prevent or detect such spoofing attacks has emerged. This thesis is concerned with so called liveness detection, that is the process of determining whether a captured fingerprint is fake or not. More precisely, the thesis explores different ways to assess how difficult it is to correctly classify a set of fake fingerprint images. Differences in image characteristics between the two classes are also explored. The purpose of the thesis is to design a quality assessment tool for fake fingerprint images used in the liveness algorithm development at Fingerprint Cards. The quality assessment tool aims to give an indication of how difficult a set of such ’spoof’ images are to classify based on the evaluated liveness characteristics.

In the first part of the thesis, features which differ between images of genuine and fake fingerprints are designed. Based on these designed liveness features, a support vector machine classifier is created by identifying the hyperplane model which best separates the images of living and spoof fingerprints. The quality of a spoof image data set is defined as the number of spoof images that this hyperplane model manages to classify correctly. Further, the quality of each individual spoof image is defined as the liveness probability assigned by the hyperplane model. Promising results were obtained from the quality assessment tool developed in the first part of the thesis. The spoof images that were assigned a low quality by the hyperplane model were images which easily could be differentiated from their live equivalents in a manual inspection. Conversely, the spoof images that were assigned a high quality were images in which the fingerprint patterns could not be differentiated from live fingerprint patterns. Hence, these results indicate a successfully designed spoof quality assessment. Further, it shows that manually designed liveness features can be used to estimate the spoof image quality.

 

In the second part of the thesis, a deep fine-tuned convolutional neural network is evaluated for quality assessment of spoof images. The utilized network has recently obtained state-of-the-art results in fingerprint liveness detection. If the deep neural network cannot differentiate between the live and fake images in a set, the images are considered very hard to classify. Conversely, if a shallow network easily differentiates spoof images from live images, these images are considered easy to classify. The liveness classification results obtained in the second part of the thesis were far better than expected. The fine-tuned convolutional neural network demonstrated fantastic liveness classification results by classifying all images in the test set correctly. These results imply that all the images in the set are possible to classify properly. However, the fact that the network managed to classify even the most realistic spoof images correctly with a high degree of certainty makes this network architecture unsuitable for spoof quality assessment. Differentiation between the images in the set could however possibly be obtained with a more shallow network.