Wednesday, April 12, 2017

Another Way To Fool The Facial Recognition Algos

Thanks (I think, as long as I don't have nightmares) to a reader.
A quick refresher:*rejFwK35SDEFuq3g.

The algorithms measure various points on the face captured by a camera and compare to known/identified images. Much of the focus is on the eyes: distance between, relationship to other features (apparent cheekbones, philtrum etc), depth of sockets and to other landmarks of facial topology.
Here's an early approach to confusing the cameras:

It is based on the distorting effects found in WWII "Dazzle" camouflage:

The next step isn't so much distorting the apparent nodal points as it is overwhelming the algos with what they think are thousands of facial "hits":

Print that on a shirt and matching billed cap and you end up with some very confused computer programs.

Finally, there's this, from Oddity Central:

Talented Makeup Artist Takes Facial Optical Illusions to a Whole New Level
31-year-old Mimi Choi, a makeup artist from Vancouver, spends hours turning her face into mind-boggling optical illusions that look photoshopped at first glance.
A former schoolteacher, Choi got into makeup only three years ago, attending classes at Blanche Macdonald, a local beauty school, to learn the basics of the craft. She’s come a long way since then, though, and today she uses her makeup skills to turn her own face into incredible optical illusions.

“When I do illusions now, I draw my inspiration mostly from my surroundings, photography, paintings, and emotions. I try not to look at other makeup artists’s work too much and challenge myself to produce original, unique work,” Choi told Allure Magazine. “My main goal each time I do a new look is to beat myself from yesterday because I’m the biggest critic and competition to myself. It’s gotten really hard to impress myself these days but it motivates me to keep trying and doing better.”...MORE
Combined with a man-bun I think the laces and bow would work for me.

Finally, here's a completely different approach via MIT: "Adversarial Images, Or How To Fool Machine Vision".