So all they would need to do is run an object detection API to put the box on the car, and only then apply OCR. Car detection is probably 99% accurate. They can do it with off the shelf implementations of Faster R-CNN or other architectures. Not to mention that they wouldn't be looking for plate numbers on the sidewalk, they would point surveillance to the street.
They don't even have to go that far... Just query the (existing) database of plate numbers and throw out anything that isn't in there. Or only run the detection for objects moving >15mph.
And even without those countermeasures, the junk data hampers mass surveillance how? It's not really preventing the scanners from reading valid plates.
How would these printed designs confuse license readers anymore than other clothes with 6-8 char text printed on them, like "ADIDAS" or "YEEZYV2"?
Also, why would these readers be scanning sidewalks and people for license plates? It seems like it wouldn't be that difficult to have a $car or $not_car image classifier.
Kind of the opposite of this though: "The worlds uggliest sweater" was supposedly programmed into the face recognition systems and would make the wearer invisable to the cameras. A deliberate backdoor.
7 comments
[ 1.8 ms ] story [ 13.0 ms ] threadHere is link to the recent Defcon presentation: https://adversarialfashion.com/pages/defcon-27-crypto-privac...
So all they would need to do is run an object detection API to put the box on the car, and only then apply OCR. Car detection is probably 99% accurate. They can do it with off the shelf implementations of Faster R-CNN or other architectures. Not to mention that they wouldn't be looking for plate numbers on the sidewalk, they would point surveillance to the street.
And even without those countermeasures, the junk data hampers mass surveillance how? It's not really preventing the scanners from reading valid plates.
Also, why would these readers be scanning sidewalks and people for license plates? It seems like it wouldn't be that difficult to have a $car or $not_car image classifier.