Master Face Attacks on Face Recognition Systems

Face authentication is now widely used, especially on mobile devices, rather than using a personal identification number or an unlock pattern, due to its convenience. It has thus become tempting target for attackers presentation attack. Traditional attacks use facial images videos of the victim. Previous work proven existence master faces, i.e., faces that match multiple enrolled templates in face recognition systems, and their extends ability attacks. In this paper, we report extensive study latent variable evolution (LVE), method commonly used generate faces. An LVE algorithm was run under various scenarios with more one database and/or system identify properties clarify which conditions strong can be generated. On basis analysis, hypothesize originate dense areas embedding spaces systems. Last but not least, simulated generated generally preserved false matching original digital forms, demonstrating poses actual threat.