Fingerprint Spoof Detector Generalization

Abstract

We present a style-transfer based wrapper, called Universal Material Generator (UMG), to improve the generalization performance of any fingerprint spoof detector against spoofs made from materials not seen during training. Specifically, we transfer the style (texture) characteristics between fingerprint images of known materials with the goal of synthesizing fingerprint images corresponding to unknown materials, that may occupy the space between the known materials in the deep feature space. Synthetic live fingerprint images are also added to the training dataset to force the CNN to learn generative-noise invariant features which discriminate between lives and spoofs. The proposed approach is shown to improve the generalization performance of a state-of-the-art spoof detector, namely Fingerprint Spoof Buster, from TDR of 75.24% to 91.78% @ FDR = 0.2%. These results are based on a large-scale dataset of 5,743 live and 4,912 spoof images fabricated using 12 different materials. Additionally, the UMG wrapper is shown to improve the average cross-sensor spoof detection performance from 67.60% to 80.63% when tested on the LivDet 2017 dataset. Training the UMG wrapper requires only 100 live fingerprint images from the target sensor, alleviating the time and resources required to generate large-scale live and spoof datasets for a new sensor. We also fabricate physical spoof artifacts using a mixture of known spoof materials to explore the role of cross-material style transfer in improving generalization performance.

Fingerprint Presentation Attack Detection Based on Local Features Encoding for Unknown Attacks

“…However, it should be noted that this method does require some samples (i.e., five) of the analysed unknown PAI species. Finally, by assuming that unknown PAIs species share texture (style) information with known PAIs, Chugh and Jain [17] extended the work in [16] by combining texture styles of pre-defined PAI species to generate new synthetic unknown PAIs. Those synthetic data could, in turn, be employed as training to enhance the generalisation capability of any end-to-end PAD approach.…”

“…Content may change prior to final publication. [17] ‡ The overall classification errors reported are the complement of the overall accuracy achieved in [44] The ACER results for the encoding fusion was attained at K = 2048 for FV and BoW, and K = 1024 for VLAD. + α = 0.9 and β = 0.1.…”

“…Chugh & Jain (FSB-v2) [13] MobileNet trained for classifying ACER = 1.11% ACER = 2.93% ACER = 14.59% ACER = 17.91% minutiae-centred local patches Chugh & Jain [14] Selection of the most feasible PAI species -BPCER500 = 24.76% -to cover the deep feature space Engelsma & Jain [15] Ensemble of several GANs -BPCER500 = 50.20% --Gajawada et al [16] Universal Material Translator (UMT) to generate -BPCER1000 = 21.96% -synthetic PAIs, thereby improving PAD generalisation Chugh & Jain (FSG) [17] Universal Material Generator (UMG)…”

Towards Fingerprint Spoofing Detection in the Terahertz Range

“…It has been shown that the selection of spoof materials used during the training process influences the performance against unknown spoofs [ 14 , 15 ]. Chugh and Jain [ 16 ] proposed a style-transfer-based method to improve the cross-material and cross-sensor generalization performance of fingerprint spoof detectors. Chugh et al [ 11 ] proposed a CNN-based method trained on patches around minutiae points which achieved impressive average accuracy of 98.61%.…”

“…Zhang et al proposed Slim-ResCNN: a deep residual convolutional neural network for fingerprint liveness detection [ 17 ], a new CNN architecture with residual blocks that achieved an overall accuracy of 95.25%. Chugh and Jain [ 18 ] proposed a style-transfer-based wrapper, called the Universal Material Generator (UMG), to improve generalization performance of any fingerprint spoof detector. The proposed approach was shown to improve performance of the Fingerprint SpoofBuster [ 19 ] and Slim-ResCNN [ 17 ] methods to a true detection rate (TDR) of 91.78% and 90.63%, respectively, at a false detection rate (FDR) of 0.2%.…”