Digital Camouflage and AI Detection
Um, the green box is, um, showing that I'm getting detected as a
person. Um, this is the, uh, YOLO algorithm running behind.
And once I'm wearing the T-shirt, I'm
not getting detected anymore as a, as a person.
Mm-hmm.
So I can basically
do like this. It's like a-
... very direct
response by the algorithm.
Mm-hmm.
My name is Simon Weckert. I'm a new media artist based in
Berlin, and mostly I work with, uh, code and electronics.
Mm-hmm.
It's called digital camouflage, yeah, yeah.
But it goes into the direction of adversary fashion.
We also see that, for example, cars, um, have a certain painting
to be invisible against flock cameras, for example, and I kind of do the
same for, um, person detection. Um, so the,
um, adversary pattern, uh, which I create is printed on a, on a
T-shirt, which you can wear, and then once you
are standing in front of an
object detection algorithm, for example, YOLO in my project, you
are getting, um, invisible. I had a pattern, like a random pattern,
and then, um, I started with, um, asking YOLO, "How
much percentage
of a person do you see when
this pattern is, uh, on the person?" And then YOLO reported like, "I see
one hundred percent," or something.
And then what the algorithm, um, the training process basically is doing
is, like, it's changing slowly patterns, colors and shapes,
flip them, rotate them, and so on and so forth, and every time after I did
that, I ask again, "How much percentage do you see now?
How much?" And so on and so forth.
And then s- after a while I realized, okay, basically the confidence
score of the algorithm went away from the category of
the person. In Berlin right now we have the situation
that it's the
first time that they're gonna use AI for camera, um, um, these
camera systems to detect, um,
most of it human behavior. You know, it's one thing that somebody's behind
a camera and watching at you, but now you have an AI, and if it
comes between the question of security and freedom, I always choose freedom,
you know? So
End of transcript · 373 words