Computer Vision News - April 2018

Last year’s competition focused on the following issues: • the presence of different operators for creating the fake fingerprints: one group for the training set and a different one for the test set; • different materials for the training set and the test set (never-seen- before materials); • the presence of some users in both the training and testing parts of the three datasets in order to explore if user-specific information can improve the system's performance. The data sets were captured by three electronic fingerprint sensors adopted to acquire more than 12,000 images: Green Bit DactyScan84C, Digital Persona U.are.U 5160 and Orcanthus Certis2 Image. The final winner of the competition, Hangzhou Jinglianwen Technology Co. Ltd , used a CNN-based algorithm to reach the average accuracy of 95.25% over the three data sets. Performance difference was noticeably lower when skilled people were involved in the presentation attack (when averaging all participants, overall accuracy dropped to 91.35%). On the other hand, if integrated in real fingerprint verification, the addition of some live and fake samples of expected users is of great help. In general, the performance achieved in this competition showed a significant improvement with respect to previous ones (see graph below). This competition was made possible through the sponsorship of Italian company Greenbit. 23 Challenge: LivDet Computer Vision News Challenge “… assess the performance of fingerprint liveness detection algorithms by using common experimental protocol and data sets. ” Average accuracy over the three latest LivDet editions The staff (l. to r.): standing Alessandra Sibiriu, Roberto Casula, Gian Luca Marcialis, Marco Garau, Valerio Mura, Marco Micheletto, Luca Ghiani, Mikel Zurutuza; sitting Giulia Orrù, Pierluigi Tuveri

RkJQdWJsaXNoZXIy NTc3NzU=