Attacking NIST biometric image software using nonlinear optimization

Sunny Raj, Jodh S. Pannu, Steven Lawrence Fernandes, Arvind Ramanathan, Laura L. Pullum, Sumit Kumar Jha

Abstract: Highlights • Attack the BOZORTH3 fingerprint matching system with 100% probability on FVC2002, ATVS and CASIA datasets. • Changing only 15 minutiae features allows us to successfully attack the BOZORTH3 fingerprint matching system. • Our approach results in better synthetic fingerprint templates compared to hill climbing technique. Abstract Automated fingerprint identification systems are deployed by law enforcement agencies all over the world for authentication. In the US, the NIST biometric image software (NBIS) is used by the Department of Homeland Security and the Federal Bureau of Investigation for fingerprint matching. NBIS uses MINDTCT as the minutia extractor and BOZORTH3 as the fingerprint matcher. We use nonlinear optimization to attack the BOZORTH3 fingerprint matching system. We use FVC2002, ATVS and CASIA datasets to validate the performance of our attack. We show that the average match score of attack fingerprints is 111.2 for FVC2002, 97.17 for ATVS and 111.07 for the CASIA dataset. We show that for all three datasets, changing only 14 minutia features allows us to attack the BOZORTH3 fingerprint matcher with more than 75% probability of successful attack.