CarCrashApocalypse — Not Quite Yet

“HACKED” STOP-SIGN CAUSES CAR CRASH — the headline pretty much writes itself, correct? The spate of self-driving cars in accidents (•), combined with all the articles about how to mess with image recognitionmeans that it’s not going to be all that far in the future when somebody will hack that stop-sign, and #CarCrashApocalypse will be here.

Or maybe not. It turns out that the story is quite a bit more complicated than that, for a very very good reason — humans already suck at seeing things.

Motorcycles
Take motorcycles for example — there is a reason why cars run into them all the time, and it isn’t because the motorcyclist did something stupid.

 /via https://www.roadandtrack.com/car-culture/a28306/why-you-dont-see-motorcycles-on-the-road/

motorcycles fall into that category of things that we don’t always perceive even if they are right in our field of vision. A motorcycle approaching head-on from a distance occupies a very small part of a driver’s vision. If it’s going quickly, it’s possible that the eye simply won’t get around to looking at it enough to make it “stick” in the brain before it arrives in the driver’s immediate vicinity. That part is important because the brain can really only see things that it understands.

Your brain has a sort of visual shorthand for objects. For instance, chances are that you aren’t really seeing everything around you right now, especially if you are in a familiar environment. You’re just seeing the shortcuts that your brain is placing there to conserve processing power and attention. That’s why people become fatigued more easily in foreign countries or really unfamiliar terrain; their brain is working overtime trying to account for all the things that it doesn’t normally see. For this same reason, if you don’t expect to see a motorcycle or pedestrian during a certain part of your morning commute, your brain will often ignore a motorcycle or pedestrian right in front of you, particularly if they aren’t moving sideways across your field of vision.

The point here is that these motorcycle/car accidents happen because car-drivers are quite bad at actually seeing. It’s not just because of selective attention, it’s because, well, it’s just the way the human eye works!

Detectors vs Classifiers
The big issue here is between Detectors (“Is that a thing?”) and Classifiers (“Is that thing a Motorcycle?”). Once a human driver actually Detects the object, they know that it’s a motorcycle, and not a school-bus, for example.

/via https://arxiv.org/pdf/1712.09665.pdf

In contrast to this, Image Recognition algorithms are ridiculously good at Detecting things. They know that there is something there — be it a motorcycle or a school-bus. All the issues that have been brought up in the media thus far have been around Classification (e.g. “is that a banana, or a toaster?”). In fact, research (••) shows that fooling the Detection process is ridiculously hard

An adversarial pattern on a physical object that could fool a detector would have to be adversarial in the face of a wide family of parametric distortions (scale; view angle; box shift inside the detector; illumination; and soon). Such a pattern would be of great theoretical and practical interest. There is currently no evidence that such patterns exist.

In summary, Humans are bad at Detecting and good at Classifying, AI is the opposite.

And that, my friends, is why we don’t actually have that much of a problem here. The car’s AI will know that there is something out there. What that thing might be up for debate, but you can add in a bunch of fairly straightforward rules to help out. Things like
• It doesn’t matter what it is, if it is in the middle of the freeway, slow down, and then stop
• A sign saying “60 mph” in a residential neighborhood is clearly wrong
• If you are wondering if it is a Street Sign or a Banana, go with Street Sign
etc.

So yeah, you really don’t need to worry about #CarCrashApocalypse quite yet!

(•) Ok, not really a spate. It’s really the opposite of spate, and, thus far, has largely been because of idiot human drivers (and yes, I’m totally ignoring anything happening at Tesla…)
(••) “Standard detectors aren’t (currently) fooled by physical adversarial stop signs” — by Jiajun Lu et al.