Hong Kong protesters started using camouflage to stay anonymous after police started using facial recognition camera's: it turns out, it is only useful if you can also hide your camouflage. They needed to be able to blend in to avoid being caught. The idea seems great, but even if this would be done for real it would not be good enough unfortunately.
post
This will not work. AI-enabled surveillance cameras are more than capable of determining that it's a wrap on a car, rather than multiple cars all somehow moving in the exact same direction in the exact same area while all facing in incorrect directions relative to the camera.
If anything, it's just another data point, because Flock cameras also catalog details about your vehicle like stickers, color, wraps, graffiti, dents, etc.
There was a story the other day where certain patterns were confusing the image detection though.
Some patterns can. This one won't.
Adversarial patterns/noise can confuse cameras, but that's usually done not by tricking it into thinking, for example, that this one van is actually a hundred mini vans crammed into the same space, but by tricking it into thinking there's not a car there because the pattern just confuses it and it recognizes it as any random sequence of random objects and things simultaneously.
Just slapping a bunch of car photos won't make it think it's a car, because the little photos are tiny, don't move like a car, are still visually distinct from the background, and are on a much larger car-shaped object. If your car has patterns that makes it look like the general concept of an asphalt road, road signs, lamp posts, etc, then it might be enough to confuse the camera into thinking the odd blob shapes are something else, because cars on the road are also always passing on top, in front of, and by asphalt, road signs, lamp posts, etc.
The cameras don't have depth perception, so if your car looks close enough to the road and objects around it, it'll blend right in, to the camera.
Wonder if covering in a design that looks like you are covered in plates would do anything. Like blurred rectangles that sorta look like a particular states plate and then random numbers. Is it illegal to cover your car in numbers?
It’s a satire account btw
The fact this needs to be pointed out shows the level of the users here.
Select all the images that have cars
Don't do this to your vehicles. A) it doesn't fucking work. B) It's meant to trick you into making yourself an easily identifiable target for law enforcement.
The best thing you can do is avoid the cameras when possible and make your car as bland as possible (no stickers, no identifying features). These systems suck, lean into that fact. Also a little "accidental" splattered mud on the license plate doesn't hurt either. Not enough to prevent a person from reading it, but just enough to break the AIs brain.
Here are some people at DEFCON who have been making low-cost devices for identifying the radio fingerprint of nearby flock cameras: https://youtube.com/watch?v=-2uAsJ5EPAw
A) it doesn’t fucking work
Have you done actual testing against multiple cam types, including Ring and Flock?
This guy has, he presented at BlackHat. And yes, it works
Plates that can't be read by software gets sorted for human input, splattering it doesnt help unless its completely unreadable.
Get an electromagnet and put behind your plates instead with a fake leaf that disappears anytime you get pulled over
Knowing that the state that submits the image for human review is that the system must automatically (1) identify a car (2) whose license plate can't be automatically read, the more productive trickery would be to defeat the "identify a car" part where the software doesn't even know that a car drive within its field of view, and therefore doesn't know to take a still frame and forward it to a human.
Alternatively, have the software automatically read a license plate incorrectly so that it doesn't trigger human review but is instead confidently incorrect in what it did log to the records.
Pretty sure flock collects a ton of data that is fed to ML algorithms such that even with a partial plate they can cross reference make/model/location/color/previous tracks to identify whose car it is or at least identify it as unique from every other vehicle.
What about those UV lights on the tag? I've wondered about using those
No way that Flock cameras have good enough resolution to be able to tell that those are pictures of cars. Not even close.
The car version of the ahegao hoodies we all wear to fool facial recognition.
It's great to see little Bobby Tables' car wrap business taking off. I need to make an appointment!
I mean can you be punished for this? It should be protected under 1st amendment imo
Maybe? Probably depends how specific the laws are around fake plates. If it's, "Displaying number plates on your car that aren't registered to it", then that's probably broad enough they could.
Instead of using a car wrap, what about just setting up a big TV on a pole, pointed directly at a camera, set to continuously generate and display images of vehicles with random license plates? In that case you wouldn't be actually displaying alternate plate numbers on any vehicle. You would just be fuzzing the camera with images of fake cars.
The entire legal justification for flock cameras is that it's legal to record anything that's in public. But it's also perfectly legal to display images of cars on a TV.
Hell, you could probably do this as a protest, holding up a TV on a pole while standing on a public sidewalk. (Might need to make it a cart of some sort.)
Yeah, I was about to say, there's no Monocacy County in Maryland :)
At the very least it would be a statement on how sick of being surveilled we are.
I wonder what would be the effect of wrapping the car with holographic paper on those flock cameras.
i doubt that would work, just knowing what i know about stuff like opencv let alone proprietary things like cognex/precise.
There are patterns that these cameras cannot parse.
They either discard the image or flag it for a human, depending on the configuration.
The reason this works is that these AI bots are trained on solid color cars and car shapes.
Eventually this might stop working, if only because these models are constantly being trained in more and more of our data.
As silly as this is, with access to the exact model flock uses for vehicle recognition, it might be possible to design a camouflage paint job that confuses the classification pipeline

top 70 comments