Thursday, November 02, 2017

iPhone X Facial Recognition (from Garret)

Garret is much smarter than I am, so here's his correction to the post I wrote on Tuesday.

Some of your assertions are not strictly true. It isn't impossible for Apple to come up with the 1 : million number, and it's not even impossible for them to test that number.

The thing about the probability is that it is likely an algorithmic calculation based on the probability of collisions in their image hash. It's not a real-world tested number, but it is likely generated by taking a massive database of sample images that are run through the facial recognition software and then checked for collisions - "matches" that are invalid. If they take 1000 images and run them through their software and 2 images match - that gives you a 2 / 1,000,000 probability that any two random faces will be erroneously matched by their software.

In fact, you would need 1,414 unique images to generate a testable 1 / 1,000,000 probability.

Now what will happen is that facial recognition will be too restrictive to start with. Different lighting conditions, different angles... it's a hard problem. If people complain that its too hard to unlock their phone because facial recognition isn't reliable enough, if it needs optimal conditions to work, then what is likely to happen is an adjustment to the algorithm to loosen those constraints. But this is a sliding scale, the more you loosen those constraints, the more likely collisions will occur. It's a balancing act between convenience and security - and that's a juggling act that will always have SOMEONE complaining about.

There are many reasons I don't like the idea of using facial recognition to unlock a device, but I won't get into them here. Accessibility wise there might be a use case where it makes sense, though I don't know what that use case is. For the general public though, I don't consider this a net positive feature.

Then, another e-mail an hour later (and just as interesting).

There are a couple of confounding factors that may cause the actual incidence of collisions being higher than the statistical incidence of collisions.

Consider a random (representative) sampling of test data, and select 1500 images. If you get a single collision, that would translate into slightly less than 1 collision per million comparisons.

But the distribution of people who look at your phone is not random, nor representative. Consider all the people in the world, which of those people is going to be mostly likely to be misidentified as you by an algorithm? The answer: your closest blood relatives. Your brothers and sisters especially (due to aging factors). What this means is that the people around you that are most likely to mistakenly unlock your phone are also the ones closest to you.

Now consider this. The chance that any random person has an identical twin is roughly 1 in 150. So if you consider a geographic sampling of people as opposed to a purely random sampling, the actual likelihood of a false-positive match is 1 in 45,000 simply due to the existence of identical twins. This is a problem that isn't likely to come up in a random image sample due to the astronomical odds of selecting BOTH twins (which is somewhere in the range of 1 in 363 billion that a sample of 1500 single portrait photos will contain pictures of each twin).

So the likelihood of your phone being able to be unlocked by another person is actually MUCH higher than 1 in a million (it's actually somewhat more frequent than 1 in 45,000).

But the numbers that Apple is giving you (although it is not explicitly stated since as far as I know they haven't published their methodology) isn't the probability that your brother or sister can unlock your phone. What they are telling you is that if your phone is stolen by a stranger, the odds of that stranger being able to unlock your phone because their face is a collision with yours is 1 in a million.

Site Meter