Deep triplet embedding representations for liveness detection

Abstract

Liveness detection is a fundamental element for all biometric systems that have to be safe against spoofing attacks at sensor level. In particular, for an attacker it is relatively easy to build a fake replica of a legitimate finger and apply it directly to the sensor, thereby fooling the system by declaring its corresponding identity. In order to ensure that the declared identity is genuine and it corresponds to the individual present at the time of capture, the task is usually formulated as a binary classification problem, where a classifier is trained to detect whether the fingerprint at hand is real or an artificial replica. In this chapter, unlike the binary classification model, a metric learning approach based on triplet convolutional networks is proposed. A representation of the fingerprint images is generated, where the distance between feature points reflects how much the real fingerprints are dissimilar from the ones generated artificially. A variant of the triplet objective function is employed, that considers patches taken from two real fingerprint and a replica (or two replicas and a real fingerprint), and gives a high penalty if the distance between the matching couple is greater than the mismatched one. Given a test fingerprint image, its liveness is established by matching its patches against a set of reference genuine and fake patches taken from the training set. The use of small networks along with a patch-based representation allows the system to perform the acquisitions in real time and provide state-of-the-art performances. Experiments are presented on several benchmark datasets for liveness detection including different sensors and fake fingerprint materials. The approach is validated on the cross-sensor and cross-material scenarios, to understand how well it generalizes to new acquisition devices, and how robust it is to new presentation attacks.