Attacks on Autonomous Vehicles Could Produce Faulty Object Locations

Researchers from Duke University’s Pratt School of Engineering have demonstrated an attack strategy that has potentially disastrous consequences for those with autonomous vehicles. The attack tricks the vehicle sensors into believing that objects are closer or further away than they actually are without being detected. The latest research suggests adding additional optical 3D capabilities or sharing data between nearby vehicles could help mitigate the problem.

All autonomous vehicles employ location/identification technology that combines 2D camera data and 3D LiDAR, which is essentially a laser-based RADAR system. They allow the vehicle to “see” objects and other vehicles around them for safe navigation on streets and highways. The combination has proven effective against malicious attacks that would otherwise cripple the navigation system by altering the area around them.

“Our goal is to understand the limitations of existing systems so that we can protect against attacks,” states Associate Professor Miroslav Pajic. “This research shows how adding just a few data points in the 3D point cloud ahead or behind of where an object actually is, can confuse these systems into making dangerous decisions.”

To confuse the autonomous system, the engineers fired a laser into the vehicle’s LiDAR sensor to add false data points. If those data points don’t coincide with the 2D information, the vehicle concludes it’s under attack and acts accordingly. Using the laser to add carefully placed data points within the 2D camera’s field of view can fool the system. The vulnerable area stretches out in front of the camera’s lens in the shape of a frustum – a 3D pyramid with its apex sliced off. A forward-facing camera will see these new data points placed in front or directly behind another nearby car and shift the vehicle’s perception by a few meters.

The team feel there is minimal risk of people setting up lasers on a car or roadside area to trick motorists passing by; however that risk increases in military applications where vehicles pass by high-value targets. The problem could become compounded if hackers manage to find a way of adding multiple false data points virtually, which could potentially attack groups of autonomous vehicles simultaneously. The researchers state that the problem could be overcome in two ways – by adding additional 2D cameras with overlapping FOVs for increased data or by sharing data with other nearby autonomous vehicles to compare data.