Page 3 of 3
Does it work?
The real question is does faceprint indentification work?
The answer is probably no.
Although there are no figures for the airline installations, a previous well-publicised trial with the Florida police force seems to demonstrate that there are big problems.
The system is impressive - it can match millions of faces per second - but it is still up to a human operator to confirm a suggested match as accurate. After two years they have discontinued the project without it helping them to catch a single criminal. It didn’t even alert them to the presence of a valid suspect, what it did do was detect lots of false positives.
The problem is that faceprint identification when used in controlled settings, i.e. ideal conditions, has a false acceptance rate (FAR) of no better than one in 250. In a real-time situation the FAR is bound to be much worse and the nature of the problem needs a much lower FAR. For example, terrorists are rare compared to ordinary people, say 1 in a million, and so to detect a single terrorist using a system with a FAR of 1 in 250 you would generate 4000 false alarms!
Even though this sounds depressing, airline companies are claiming to be making flying more secure by scanning for terrorists using face recognition. As the systems are based on the more sophisticated EigenFace it might just be that they work better – but given the very high accuracies needed to detect rare events, such as a terrorist boarding a plane, this seems unlikely. It is argued that inefficient security provides false security and other systems should be used.
Things aren’t all gloomy for face recognition however. The error rate is good enough to allow face recognition to be used as a backup to other forms of identity checks and the techniques are getting better all the time.
Face recognition has still got a long way to go before it lives up to the dreams of recognising the criminal or terrorist using CTV surveillance cameras. It has a much better chance of working well if you can improve the quality of the input by getting people to co-operate and stand in front of a camera. This is fine when people want to be correctly recognised, such as in a PC or credit card validation system.
It’s not so good when they might be trying to avoid being recognised as in the terrorist detection system where a good disguise currently foils most recognition systems including humans!