Human-driven cars can also be disabled with a laser pointer to the eye of the driver.
"Attacks" on self-driving cars that could already have been conducted on humans, and clearly aren't, because it just isn't worth anybody's time, aren't very technically interesting. One could make a case that there's a socially interesting aspect in that someone could use such an attack, in conjunction with social fears of self-driving cars, to do something like trash somebody's stock price or something, but the "social fear of self-driving cars" would be an irreducible element of such an attack.
It's the attacks like the Jeep attack, where I can hit fleets at a time, because the cars are all intimately networked together or something, that are interesting.
We really need to remember that humans are atrocious at driving on average, and doing better than them is a pretty low bar. Not that we shouldn't strive to improve automation, but don't let perfect be the enemy of the good.
Let's revisit that assertion when autonomous vehicles manage to even minimally cope with heavy rain or snow, neither of which are barriers for those "atrocious" human drivers.
The reality is that, for all their flaws, humans vastly outperform autonomous vehicles in imperfect driving conditions. That's just reality, despite the HN tendency to pitch this technology as though it could replace all human drivers tomorrow.
My point is simply that there are tasks where human drivers outperform computers, and there are tasks where computers outperform human drivers, so simply looking at some scenario where humans do better doesn't tell us anything. You have to look at the whole picture.
And yes, currently humans are way better than computers in the general case. I'm not disputing that. I'm just saying that "computers can't handle situation X, which humans can" is a pointless thing to say by itself.
ROFL, no, with absolutely no qualification, you flat out said:
We really need to remember that humans are atrocious at driving on average, and doing better than them is a pretty low bar.
I guess what you're saying is that computers are so bad, they still can't get over that "pretty low bar"?
If so, I suppose I agree, though I think you're dramatically underestimating just how capable human beings are.
so simply looking at some scenario where humans do better doesn't tell us anything.
Which is really strange, because that's exactly what everyone does when talking about autonomous vehicles. The discussion always surrounds current trials, where vehicles are allowed to operate in cherrypicked, near-ideal conditions: perfect, southern California climate, with minimal rain, zero snow, and perfect visibility.
So I agree, let's stop looking at just some isolated scenario. Let's focus on the general case... where computers, unfortunately, currently fail, and will likely continue to do so for a very long time (dealing with highly variable conditions likely requires major infrastructure changes to actively support autonomous vehicles, and that ain't happening any time soon).
Yes, that is exactly what I'm saying. So you can skip over the parts where you talk about how other people, not me, praise the abilities of current autonomous vehicles, because I'm not doing that.
The bar is actually quite high, if you want to do better than humans on average.
And in this particular case, humans not only can be disabled with a laser pointer, but they can be disabled permanently, not just for that driving session, but for the rest of their lives. Can't swap out cameras in a human driver. Yet somehow this is not a crisis for manually-piloted cars.
I doubt it happens that often, but I can't back that up with numbers, and I can't be arsed to try.
But, humans can vary in their driving ability. In the case of autonomous cars, you would have to assume any flaw in one is shared by all. The problem isn't the laser pointers, or that humans can also be disabled by laser pointers, but that autonomous cars aren't yet smart enough to deal with errant signals. If all the autonomous cars can be stopped by an ever increasing list of simple conditions that (with a couple of exceptions, like a laser pointer in the eyes) wouldn't stop an average human driver, then it implies that the current state of the art does not nearly live up to the hype.
That's not to say that computers are superior to humans. It's quite clear that the current state of the art in autonomous driving is way behind what humans can do. My point is simply that comparing scenarios which one can handle and the other cannot is not very informative.
My guess is that either the driver's vision will be quickly impaired or his/her eyes will quickly start hurting. In that case, the driver will quickly know something is wrong even if he/she may not know it's an attack.
I'm no expert on laser safety or human biology, so I left a question on biology SE: http://biology.stackexchange.com/questions/38603/how-quickly...
Hopefully somebody can give an answer.
And if the driver experiences no symptom, then the attack is just ineffective.