Photo: Cornell University research team
For us human drivers, traffic signs don't get more straightforward than the "Stop" one. Not only is it big and red, but it also says "STOP" in white, using a clean font. When you see this one, you don't have to open the highway code and see what you're supposed to do. You push the brakes and, well, stop.
For autonomous cars, though, things aren't as simple as that. They don't have years of their mothers telling them to stop, sometimes even yelling, so the notion is a lot more abstract. They don't have cognitive minds either, so even though they can learn, the process is a lot different to what we're used to.
Right now, traffic sign recognition relies on the vehicle's video cameras, which means the sensors capture the image and the AI decides what it is it's looking at. Once that's done (and, with the current CPUs, we're talking nanoseconds here), the vehicle takes the appropriate action.
But there's a weak link here. If, for instance, a bird takes a weird poop on a "Stop" sign, could the vehicle be fooled into thinking it's actually a speed limit indicator? A group of researchers believes it could, and that's a bit scary.
Since training birds to relieve themselves on a specific target seemed too complicated, the team at Cornell University used a few simple black and white stickers which they applied on the sign. Placed in the correct sequence, they were enough to convince the vehicle it was looking at a speed restriction in 100 percent of the cases.
A second experiment tinkered with a Right Turn sign, which the car's AI mistook for either a Stop or Added Lane sign in 100 percent of the testing conditions as well. Needless to say, this poses a problem for driverless cars, but also for the many vehicles that already use sign recognition software.
The cars don't rely solely on what they see, but also cross-check the information with the onboard database, but in the event of driverless cars, the potential outcome of such an attack could be disastrous. Not to mention it could all be carried out using nothing more than some self-adhesive paper and a printer.
Luckily, despite the strong push from automakers, the authorities are in no rush to see self-driving vehicles on the roads just yet. The technology clearly needs a few more years of testing during which all these possible security breaches will hopefully be ironed out, so we're not exactly there yet.
Right now, traffic sign recognition relies on the vehicle's video cameras, which means the sensors capture the image and the AI decides what it is it's looking at. Once that's done (and, with the current CPUs, we're talking nanoseconds here), the vehicle takes the appropriate action.
But there's a weak link here. If, for instance, a bird takes a weird poop on a "Stop" sign, could the vehicle be fooled into thinking it's actually a speed limit indicator? A group of researchers believes it could, and that's a bit scary.
Since training birds to relieve themselves on a specific target seemed too complicated, the team at Cornell University used a few simple black and white stickers which they applied on the sign. Placed in the correct sequence, they were enough to convince the vehicle it was looking at a speed restriction in 100 percent of the cases.
A second experiment tinkered with a Right Turn sign, which the car's AI mistook for either a Stop or Added Lane sign in 100 percent of the testing conditions as well. Needless to say, this poses a problem for driverless cars, but also for the many vehicles that already use sign recognition software.
The cars don't rely solely on what they see, but also cross-check the information with the onboard database, but in the event of driverless cars, the potential outcome of such an attack could be disastrous. Not to mention it could all be carried out using nothing more than some self-adhesive paper and a printer.
Luckily, despite the strong push from automakers, the authorities are in no rush to see self-driving vehicles on the roads just yet. The technology clearly needs a few more years of testing during which all these possible security breaches will hopefully be ironed out, so we're not exactly there yet.
Google News