Assessing Transfer Learning on Convolutional Neural Networks for patch--based Fingerprint Liveness Detection

Fingerprint based biometric identification systems are vulnerable to spoofing attacks that involve the use of fake replicas of real fingerprints. The resulting security issues can be mitigated through the development of software modules capable of detecting the liveness of an input image and, thus, of discarding fake fingerprints before the classification step. In this work we present a fingerprint liveness detection method that combines a patch--based voting approach with Transfer Learning techniques. Fingerprint images are first segmented to discard background information. Then, small--sized foreground patches are extracted and processed by popular Convolutional Neural Network models, whose pre--trained versions were adapted to the problem at hand. Finally, the individual patch scores are combined to obtain the fingerprint label. Experimental results on well--established benchmarks show the promising performance of the proposed method compared with several state-of-the-art algorithms.