This essay is currently incomplete and missing the middle part, where I actually do the math. In the meantime, this article does the type of math I need to include in that section. My unique contribution to the debate is after "It's Impossible". (Nobody was supposed to find this.)

Ladies and gentlemen, I'm afraid I have some bad news. By the end of reading this essay, I guarentee that one of two things will be true:

1. You will believe that facial recognition technology will never work as an airport security measure.
    
2. You ascribe magical capabilities to computers, even after repeated demonstrations of impossibility.
    

In the end, if you insist on believing in magic, that's your decision, but I intend to leave that your only choice.

By the way, in case you don't know me, you should know I'm a grad student in computer science. I'm not just making this up out of thin air, I've seen the state of the art. And I can tell you for certain that what the snake oil security software vendors are trying to sell can't possibly be working, and in fact, isn't.

There are basically two problems I've seen this technology touted as a solution for:

1. Attempting to identify people at a security checkin point, and pluck the terrorists out of the hundreds of thousands of non-terrorists the system will see.
    
2. Putting the camera up in some public location, like an airport terminal hallway, and pick the terrorists out of the feed. (This is also suggested for things like football games and concerts; the problems I outline below apply to these cases too.)
    

There are real differences between these two problems, which is why I want to mention them seperately at the beginning, but basically, the differences can be accurately summed up by saying that #2 is a lot harder then #1... #2 faces all of the problems #1 does, plus some news one of its own, like trying to pluck faces out of a generic picture, and trying to identify faces from much less data (because the pictures of each individual person are smaller). Fortunately for the purposes of this essay, #1 is safely in the domain of the impossible, so I will show #1 is totally impossible, which makes #2 even "more" totally impossible.

The math matches reality. Consider this news article about the implemented systems in Florida.

There's something that I haven't seen anybody point out, and if you still have doubt, this really ought to clinch my point. The thing is, everywhere I've said facial recognition system, it holds for any system, not just a computerized one. In fact, it hold for human beings as well.

Remember when I said the system needs 99.99999999999% accuracy, and it was hard to get any grip on that number to understand it? Let me put this another way. Have you ever in your life misidentified someone, for any reason? New glasses caused you to not recognize someone? Someone you thought was an old chum from school turns out to be someone else entirely? Light too dim, too bright, they were wearing something odd, anything? Since I'm willing to guarentee that you have not identified someone more then a billion times in your lifetime, that means that you are not good enough to be a security face recognition system with any sort of accuracy! By many factors of magnitude!

And we come full circle back to the original problem. Why do we need these computerized facial recognition systems in the first place? Because the humans we have manning the security stations aren't good enough at face recognition to do it themselves!

Sorry, but like I said before, there's no significant vision task that computers are better at then humans.

As a consequence of analysing the problem, and observing that the analysis applies to all facial recognition systems, whether they involve humans or computers, we can draw some solid conclusions. For the first problem formulation, with the ambient cameras in the hallways trying to identify people off of the equivalent of a security camera feed, we can safely say that solving the first problem is flatly impossible, no matter what technology you throw at the problem. There is simply not going to be a way to push the accuracy threshold high enough to pick a small group of people out of a huge one without getting hundreds or thousands of false positives per real positive... to say nothing of the false negatives.

It's also worth observing at this point that our problem formulation implicitly assumes that the terrorists are dumb, and not trying to get past the system. We have nearly a hundred years of experience in fooling cameras, from Al Jolson's race transformation for the final scene of "The Jazz Singer" up through whatever the latest cosmetic surgery wonder is, with makeup and disguises of varying effectiveness in between.

With your hopefully-enhanced understanding of the problem, hopefully you find this easier to believe now then before. (Just remember, can you imagine trying yourself to identify people from just a security camera video feed?)

As for the other problem we specified, we have to come to the same conclusion. In terms of plucking out a small group of people from a pool of millions, again assuming they aren't even making a half-hearted attempt to disguise themselves, the accuracy required is still absurdly high, beyond even a human being's ability.

. . . the systems you've seen on TV that seem to work?

The only face recognition systems I have ever seen work with any sort of accuracy are systems where a cooperating person sits down in front of the computer, in known, constant lighting, with a high-quality camera focused nearly directly on the face, with the human being not trying to fool the system, with a small pool of faces to recognize (tens, maybe hundreds), and an equally small pool of people who may be trying to be identified who aren't in the database (little space for false positives).

And even if these systems are used for security, they are used as only part of the identifying process; the user must still present a password or token or some other identification.

It can be difficult to understand how the problems stack up as you start tweaking each parameter... as the human is no longer cooperating, as the lighting is inconsistent, as the camera drops in quality and expands its field of view, as the pool increases to looking at millions of faces, and (worst of all) as the human starts trying to fool the system, the problems just stack up until it can't be done, just as our math shows above.

Consider your problem of staying alive. You need air (of a certain blend of oxygen and inert filler), a pressure range, a certain temperature range, food, water, gravity, and the absence of excessive radiation, chemicals, water, etc. You can tolerate some of these parameters going outside of your normal range for some period of time (too cold, too hot, underwater, no air), but the more problems you have to deal with at once, the harder it is to stay alive. Very quickly, it becomes impossible. Face recognition can be likened to our body; it absolutely requires a friendly environment to work in. Trying to make it work in an airport situation is like trying to survive in the cold vacuum of outer space without a space suit. The temperature is wrong, the pressure is wrong, there's no food, no water, excessive radiation, basically, everything is stacked against you. In that environment, survival is impossible. So is facial recognition technology.

What do you know about them now?

• You know the false positive rate is absurdly high. Thousands of false positives or more for every real positive. So absurdly high, in fact, that the best thing a human user of the system can do is simply ignore the system. The worst thing a human can do is listen to the system; the user won't catch anyone, but they will annoy (or worse!) thousands of people.
   
• You know false negatives still don't disappear.
   
• You know it's trivially easy to bypass the system if you're a dedicated terrorist.
   
• You know the necessary accuracy can never be acheived.
   
• For such a useless system, it's very, very expensive; it's wasted money better applied elsewhere, on security that has a chance of actually working.
   

In summation, the situation is better without attempted facial recognition technology then it is with is. Facial recognition technology is of negative value. If it weren't for the fact that people still trust the technology and have an guilty-until-proven-innocent trust of the computer's results, this wouldn't even be worth fighting; this fact will eventually (very eventually!) be recognized by the airports. (Were it not for the blind trust people place in computers, nobody would even be suggesting this as a solution!) But there will be real victims of this technology if anybody tries to deploy it. On that basis, facial recognition technology must not be deployed for security purposes in airports.