Tesla Motors' vehicles have been under close scrutiny after several incidents and a reported autopilot-related fatality. The government has already vowed investigations, and the reliability of the system is being questioned.
But what if somebody wanted to crash a Tesla set to autopilot? Could the system actually be fooled into crashing? As it turns out, yes.
Academics from Zhejiang University, University of South Carolina, and Chinese security giant, Qihoo 360, have been able to fool Tesla's autopilot's sensors. The researchers tricked the sensors, making them perceive an object where none existed, or overlook a real object in the Tesla’s path.
The researchers attempted to fool all three of Tesla's autopilot's detection systems, radar, ultrasonic sensors, and cameras. To fool the radar systems, they used two pieces of radio equipment—a $90,000 signal generator from Keysight Technologies and an equally-expensive VDI frequency multiplier. Both were used to drown out the radio waves bouncing from the cart back to the Tesla, making the obstacle "invisible" to Tesla’s autopilot and disappear from its screen.
Attacking the short range ultrasonic sensors proved to be easier and cheaper, yet far less fatal. These sensors are used in the self-parking and “summon” features. Using a small generator to create certain voltages, and an ultrasonic transducer to convert that electricity to sound waves (totaling no more than $40), the researchers were able to make a Tesla not park for fear of an imaginary obstacle, or miss an obstacle and park anyway.
But for the final system—the camera— they weren't so lucky. They were able to blind the cameras, but all this did is switch off autopilot and give the car back to the driver.
While the tests prove that it is possible, Wenyuan Xu, the lead researcher notes the limits of the attacks. “Only very highly motivated attackers, with budgets, will be able to launch this type of attack,” she says to Forbes. The team will be presenting the results at this year’s DEF CON conference.