
Driverless cars could be tricked into interpreting a red traffic light as green, leaving them vulnerable to attacker-instigated crashes.
The legislation around autonomous vehicles varies worldwide. Although fully driverless cars aren’t yet on the roads, with such vehicles in development, concerns have been raised that they may be more vulnerable to vandalism and hence robust regulation is needed to protect against this.
To learn more, at Zhejiang University in Hangzhou, China, and his colleagues tested whether the cameras of different driverless models could be tricked into incorrectly interpreting a traffic light.
Advertisement
The team aimed a laser at the sensors of five camera models used by driverless cars, with two open-source software packages interpreting the camera-captured images.
A laser with a 650-nanometre wavelength introduced colour errors so that the entire image turned red, while a 520-nanometre wavelength gave the image a green hue, neither of which fooled the cameras.
When flickering the laser at high frequencies, however, only certain parts of the image took on these colours. Adding a horizontal bar of green or red caused both software packages to incorrectly interpret the traffic lights as green 30 per cent of the time and red 86 per cent of the time, on average across the five cameras.
The authors didn’t respond to a request for comment, but suggest in their paper that redesigning driverless car cameras so they scan rows of pixels in a random order, rather than one row at a time, would make it harder for an attacker to introduce colour interference.
at De Montfort University in the UK says research like this highlights the need for strong regulation of driverless cars, which aren’t overseen by a global regulating body.
“What needs to happen is autonomous vehicles need to be under as stringent guidelines as the airlines and aircraft manufacturers,” he says. “If we don’t have that, it’s going to go badly, badly wrong. That’s why studies like this are so good and so important, because they help to highlight potential issues that the car manufacturers need to be thinking about now.”
The experiment required a lot of work to trick just one sensor, and driverless cars will hopefully have systems so their software doesn’t rely on one camera alone. However, a determined attacker with enough resources “will always be able to make it go wrong”, according to Smith.
Reference: