I recently renewed my passport and got one of the new ePassports which allows me to use the SmartGate passport control. This system uses facial recognition technology for self-processing at the border control.
How accurate is facial recognition techology? It is far from perfect, but it is very useful when used correctly in the right system.
Accuracy of biometric recognition systems is measured in two ways: False Acceptance Rate (FAR) where the system recognises a face when it should not have (e.g. letting an unknown person in), and False Rejection Rate (FRR) where the system does not recognise a face when it should have (e.g. keeping a legitimate person out).
These two values are related, because you can always improve one at the expense of the other. For example, tune the algorithm to accept more borderline cases and you improve the FRR but make the FAR worse. Letting everyone (including the bad guys) into the country is perfect FRR, but terrible FAR. Keeping everyone out (including legitimate people) is perfect FAR, but terrible FRR.
To see some real numbers, I found the results from the Face Recognition Vendor Test (PDF) of 2006.
Their benchmark performance for 2006 technolgy, for a FAR of 0.001, was a FRR of 0.01. That is, incorrectly accepting 1 in 1000 faces means incorrectly rejecting 1 in 100 faces. That was the benchmark, but different algorithms achieved similar or worse results under different conditions (see the graphs on pages 14 and 16). A FRR of 0.01 sounds poor, but is significantly better than the FRR of 0.79 that was achieved using 1993 technology -- incorrectly rejecting 8 out of 10 faces!
If the facial recognition algorithms are so poor why are they being used at border control? It is because it is not just the algorithm, but the entire system that counts.
In these systems, they probably use algorithms that crank up the FAR, so that the computer is very unlikely to let the wrong person into the country. That means their FRR is poor, so more legitimate people will be refused entry. But those people can then be processed by a border control officer, who can then recognise them. So it is not the algorithm that works, but the entire system involving both computers and people that works.
This system actually uses the respective strengths of people and computers. Page 20 of the report shows the error rates of people and the algorithms. It shows that when the FRR is high, the algorithms generally achieve better FAR than people; but when the FAR is high, people achieve better FRR than the algorithms. That is, an algorithm is better at correctly rejecting an impersonator than a person; but a person is more better at correctly recognising a person (even though they might look different) than a computer.
It is what we expect: computers are not very good at recognising faces. People are better at recognising faces, but computers are better at rejecting faces. Together the system works. Perhaps researchers should really be claiming success at facial rejection technology rather than facial recognition technology!