News Yous Tin Flame Use—Invisible Mask: Practical Attacks On Confront Recognition Amongst Infrared

From arXiv:

Zhe Zhou, Di Tang, Xiaofeng Wang, Weili Han, Xiangyu Liu, Kehuan Zhang

(Submitted on xiii Mar 2018)

Accurate confront recognition techniques build a serial of critical applications possible: policemen could employ it to holler back criminals' faces from surveillance video streams; cross boarder travelers could transcend a confront authentication inspection trouble without the interest of officers. Nonetheless, when world safety heavily relies on such intelligent systems, the designers should deliberately visit the emerging attacks aiming at misleading those systems employing confront recognition.

We suggest a variety of build novel ready on against confront recognition systems, which is realized past times illuminating the champaign of report using infrared according to the adversarial examples worked out past times our algorithm, so confront recognition systems tin dismiss last bypassed or misled spell simultaneously the infrared perturbations cannot last observed past times raw eyes.

Through launching this variety of attack, an assaulter non exclusively tin dismiss dodge surveillance cameras. More importantly, he tin dismiss impersonate his target victim together with transcend the confront authentication system, if exclusively the victim's photograph is acquired past times the attacker. Again, the ready on is totally unobservable past times nearby people, because non exclusively the lite is invisible, only besides the device nosotros made to launch the ready on is pocket-size enough. According to our report on a large dataset, attackers convey a really high success charge per unit of measurement amongst a over 70\% success charge per unit of measurement for finding such an adversarial event that tin dismiss last implemented past times infrared. To the best of our knowledge, our operate is the offset 1 to shed lite on the severity of threat resulted from infrared adversarial examples against confront recognition.

Subjects:	Cryptography together with Security (cs.CR)
Cite as:	arXiv:1803.04683 [cs.CR]
	(or arXiv:1803.04683v1 [cs.CR] for this version)

Submission history 

From: Zhe Zhou [view email]
[v1] Tue, xiii Mar 2018 08:51:01 GMT (3397kb,D)

Which authors of this newspaper are endorsers? | Disable MathJax (What is MathJax?)

arXive download site at Cornell