Image: University of Washington/Robust Physical-World Attacks on Machine Learning Models.
Researchers have discovered that like human drivers, automated vehicles can also get confused when encountering vandalized road signs.
A team composed of members from the University of Washington, University of Michigan Ann Arbor, Stony Brook University, and University of California Berkeley did a study on how defacing road signs can affect the artificial intelligence in self-driving cars. They published their study called “Robust Physical-World Attacks on Machine Learning Models”.
For their experiments, the researchers used stickers to alter how the artificial intelligence of the car sees a stop sign. By adding the stickers “love/hate” on a stop sign, the computer mistook it for a speed limit sign.
Human drivers may be more aware of vandals defacing traffic signs, but autonomous vehicles have no such experience in these situations. Thus, this poses a threat to the safety of passengers.
Fooling the AI is also relatively easy to do since the stickers can be printed even at home. All that’s left is for the attacker or prankster to know how the car’s AI classifies images.
The researchers believe that their findings would be helpful in creating driving systems that are more robust against future attacks. Considering that autonomous vehicles are being hailed as the future of transportation, it would ease the worries of passengers if the autonomous car one is riding can properly read road signs. JB