Use Baidu's platform to show how the fusion of Lidar, radar, and cameras can be fooled by stuff from your kids' craft box
A team of researchers from prominent universities – including SUNY Buffalo, Iowa State, UNC Charlotte, and Purdue – were able to turn an autonomous vehicle (AV) operated on the open sourced Apollo driving platform from Chinese web giant Baidu into a deadly weapon by tricking its multi-sensor fusion system, and suggest the attack could be applied to other self-driving cars.
TL;DR: faking out a self-driving system is always going to be possible, and so is faking out humans. But doing so is basically attempted murder, which is why the existence of an exploit like this is not interesting or new. You could also cut the brake lines or rig a bomb to it.
I think human responses vary too much: could you follow a strategy that makes 50% of human drivers crash reliably? probably. Could you follow a strategy to make 100% of autonomous vehicles crash reliably? Almost certainly.
You don't even have to rig a bomb, a better analogy to the sensor spoofing would be to just shine a sufficiently bright light in the driver's eyes from the opposite side of the road. Things will go sideways real quick.
It’s not meant to be a perfect example. It’s a comparable principle. Subverting the self-driving like that is more or less equivalent to any other means of attempting to kill someone with their car.
More exciting would be an exploit that renders an unmoving car useless. But exploits like this absolutely will be used in cases were tire-slashing might be used, such as harassing genocidal vips or disrupting police services, especially if it's difficult to trace the drone to its controller.
This is the real reason Elon Musk doesn't want people tracking his plane. If we know where he is, Wile E Coyote could catch up to him and trick his car into crashing into a brick wall, by painting a tunnel on it
I mean, not this certain attack, but the principle is well known.
The solution is also known: any sensor (or at least any critically important sensor) in a robotic system must be able to recognize it's own state of "blindness". The system must react accordingly. (For example, with the camera behind the windshield, it would activate the wipers and the heating in the windshield to remove possible rain, snow or dirt). If several sensors go "blind" at the same time, the system must do a safe stop of the car.
It's basically chaff, lol. We've known chaff is an effective radar countermeasure since the 40s, and it seems like the researchers have found the lidar and optical equivalents of chaff. What really scares me is the idea of this evolving into more sophisticated deception attacks like range or velocity gate pulls. No idea how you'd do that with lidar or optically, but I'd bet money that's a line item on a black budget somewhere
Its still a problem, A sheet held across the road on a string would show up as a wall to both cameras and lidar. I for one am buffalo buffalo buffalo buffalo buffalo looking forward to the emerging profession of road pirates robbing automated trucks this way.
Ok but the problem of road pirates isn't new either, is it? Let's watch 'Herbie' again :-)
There is just one risk that is kinda new (but actually coming with every automation): systematic errors could bring vulnerabilities that get exploited in large numbers.
I just want a cheap non internal combustion engine. I don't care about self driving bullshit. I have eyes, arms, legs, and a brain, I'll do the driving.