Neural network technology helps self-driving cars identify phantom objects

Researchers at the Negev Center for Cyber Security Research at Ben Gurion University in Israel say projection images on the road to form phantom objects can cause a semi-autonomous or fully autonomous car in motion to misjudge and brake sharply, endangering the lives of drivers and passengers inside the vehicle. The neural network technology they are working on will solve the problem of self-driving cars failing to recognize phantom objects.

Neural network technology helps self-driving cars identify phantom objects

The team’s demonstration showed that projecting unreal objects such as characters or unreal road routes on the road could lead to a brake or run-off of a semi-autonomous or fully autonomous vehicle. Ben Nassy, a doctoral student in the research team, believes that this behavior can be called a “phantom attack” using drones equipped with portable projectors or roadside digital billboards that break into Internet control.

Nasi also said that the remote use of drones and digital billboards to carry out phantom attacks on autonomous vehicles, usually do not leave any evidence on the scene, attackers do not need any complex preparation, while the devices used to carry out the attack are inexpensive. The team has shown that the driver’s aid system of self-driving cars can be taken to the truth by simply showing the route Phantom on a roadside digital billboard for 125 milliseconds.

Neural network technology helps self-driving cars identify phantom objects

The automotive industry has not yet considered phantom attacks. Naxi believes that the object recognition system used in self-driving cars is flawed, and that it essentially uses feature matching to detect visuals without being trained to distinguish between real and false objects. The researchers also point out that, in fact, even with depth sensors, the absence of deep objects projected on the road is considered true, as a result of the automotive industry’s “safety is better than regret” policy, which allows automotive identification technology to treat visible two-dimensional objects as real objects.

The researchers recommend that automakers take steps to ensure that autonomous driving technology is able to identify objects on the road. At the same time, researchers say they are developing a neural network to examine and detect reflected light from objects, as well as their surfaces and backgrounds, to determine whether objects on the road are real or phantom, providing solutions for automakers. (Reporter Mao Li)