WEBVTT

00:00:00.000 --> 00:00:03.360
What if I told you you could defeat incredibly high tech

00:00:03.360 --> 00:00:06.400
surveillance with a decidedly low tech tool?

00:00:06.400 --> 00:00:11.400
And no, I'm not talking about spray painting a security camera, I'm talking about simple clothing

00:00:11.400 --> 00:00:15.000
that can throw off advanced AI algorithms and facial recognition.

00:00:15.000 --> 00:00:20.360
It turns out that certain algorithms can be essentially reverse engineered in a manner that fools them

00:00:20.360 --> 00:00:25.600
even if your face is in plain view. A few years ago, researchers at a university in Belgium

00:00:25.600 --> 00:00:28.680
took a closer look at a common image recognition algorithm

00:00:28.720 --> 00:00:32.040
called YOLO V2, it's literally the name,

00:00:32.040 --> 00:00:36.960
and worked out how to produce images that the algorithm either wouldn't recognize as human

00:00:36.960 --> 00:00:40.040
or wouldn't even recognize as any kind of object.

00:00:40.040 --> 00:00:46.000
The resulting images, called adversarial patterns, often resemble brightly colored abstract art

00:00:46.000 --> 00:00:49.840
that the algorithm can't make heads or tails of. But can any of us?

00:00:49.840 --> 00:00:53.140
Not only does this lead the algorithm to think there isn't a person in the image at all,

00:00:53.140 --> 00:00:57.120
but the patterns also make for some pretty cool looking clothing that there's apparently

00:00:57.120 --> 00:01:00.540
already a market for. Some of the patterns effectively make the wearer

00:01:00.540 --> 00:01:03.820
invisible to AI, while others make the computer think

00:01:03.820 --> 00:01:08.340
that it's actually looking at something else, like a dog. Of course, the method isn't foolproof,

00:01:08.340 --> 00:01:11.620
as a more sophisticated algorithm could get released at any time

00:01:11.620 --> 00:01:16.580
and would still recognize what it's looking at, and it won't do much if an actual person

00:01:16.580 --> 00:01:22.060
is sitting on the other side of the security camera either. But why bother trying to be invisible to AI

00:01:22.060 --> 00:01:27.140
when you could just spam it with noise? We're starting to see clothing made to resemble

00:01:27.140 --> 00:01:30.300
other objects that automated cameras commonly look for,

00:01:30.300 --> 00:01:33.360
such as this dress that's covered with license plates,

00:01:33.360 --> 00:01:38.440
or this scarf with photorealistic faces. This way, you don't have to beat its algorithm,

00:01:38.440 --> 00:01:43.140
you just feed it with junk data, and you're also making a rather bold fashion statement.

00:01:43.140 --> 00:01:46.380
Of course, if you're serious about remaining hidden, you could try this approach,

00:01:46.380 --> 00:01:52.060
which uses these masks that have a lens-like effect to trick cameras, but keep you recognizable

00:01:52.060 --> 00:01:57.820
to fellow humans, or even this prototype that projects different faces onto your own.

00:01:57.820 --> 00:02:02.380
It's like a scanner, darkly, assuming that you can be bothered to wear something like that around town.

00:02:02.380 --> 00:02:05.940
But regardless of what method anyone uses to fool an AI camera system,

00:02:05.940 --> 00:02:09.340
exactly how widespread are these surveillance technologies

00:02:09.340 --> 00:02:13.340
that researchers and designers have felt the need to explore ways to beat them?

00:02:13.340 --> 00:02:17.020
Actually, more widespread than you might think. There have been plenty of headlines

00:02:17.020 --> 00:02:21.900
about how the authorities in mainland China have used smart cameras to track and identify citizens,

00:02:21.900 --> 00:02:26.500
but dozens of other countries are using similar technologies for everything from cracking down on dissidents

00:02:26.500 --> 00:02:31.940
to more benign pursuits like traffic management and smart cities, and crowd security at sporting events.

00:02:31.940 --> 00:02:37.660
And these networks largely cannot function if they can't even detect objects in the first place.

00:02:37.660 --> 00:02:42.340
Regardless of why you're being spied on, it's understandable if you want to try it and opt out

00:02:42.340 --> 00:02:46.500
while still being able to walk around a city center. But fair warning, even clothing

00:02:46.500 --> 00:02:49.900
that utilizes adversarial patterns doesn't work all the time,

00:02:49.900 --> 00:02:53.940
even against image recognition algorithms it was designed for. Like if maybe you're just standing

00:02:53.940 --> 00:02:56.940
kind of two sideways or something, you know?

00:02:56.940 --> 00:03:00.580
Suddenly setting up my own society in the woods sounds pretty appealing.

00:03:00.580 --> 00:03:04.300
Thanks for watching guys, if you liked this video, hit like, hit subscribe, and hit us up in the comments section

00:03:04.300 --> 00:03:07.300
with your suggestions for topics that we should cover in the future.