Robust and interoperable fingerprint spoof detection via convolutional neural networks

Fingerprint recognition for automated border control and other high-security applications needs robust integrated anti-spoofing capability. Facing the threat of presentation attacks, two key challenges to be solved are sensor interoperability and robustness versus new fabrication materials. This paper proposes convolutional neural networks for this task and presents an exhaustive comparison on latest LivDet 2011 and 2013 databases. Apart from classical classification nets, also metric-based deep siamese networks are evaluated learning a distance metric enforcing live-spoof pairs to be of higher distance than live-live pairs. This is useful for attended enrollment scenarios where a live gallery image is available (e.g. trusted-source fingerprint reference on the passport chip). Experiments reveal remarkable accuracy for all Convolutional Neural Networks (CNNs) CaffeNet (96.5%), GoogLeNet (96.6%), Siamese (93.1%), good material robustness (max. 5.6% diff.) but weak sensor-interoperability.