Self-driving cars are statistically and logically far less likely to do something like speed or run a stop sign, but researchers have found that defacing street signs can change all of that. A recent study found that while such vehicles could always figure out that they were looking at a street sign of some sort, it was pretty easy to convince them that they were looking at a different sign than what was actually there. Since street signs get vandalized all the time in the real world, the implications here are rather clear and possibly alarming. Addressing the problem will likely involve feeding the cars large sets of data on how to identify street signs that have been defaced, and what to do if an unreadable or unidentifiable street sign is encountered. No self-driving car makers have issued any official comment on this issue as of this writing.
During the course of the study, researchers from various universities teamed up to discover a few things about how self-driving cars worked and how best to confuse them. Examples included putting stickers on signs, creating fake signs with small differences from the real ones, and even embedding attack algorithms into patterns that can be overlaid on a sign and look fairly normal to the human eye. Most of those examples focused on getting a self-driving car to identify any given sign as a speed limit sign with a speed limit of 45, and they succeeded. Some attacks were also focused on making the systems identify any given sign as a stop sign, and these were also successful. The stickers used were printed on a normal, consumer-grade printer.
The white paper did not say whose self-driving car systems the study used, or if the systems were made for the study and simply patterned after existing solutions. Regardless, the study proves that given the current state of this emerging technology, it’s entirely feasible for somebody to simply create and print a rash of fake street signs and sign decorations in order to wreak havoc among autonomous vehicles, the implications of which should be relatively obvious, and it remains to be seen whether the industry manages to come up with a method of addressing this issue in the immediate future.
