Sunday, July 15, 2007

CCTV Face Recognition: Not Just Evil, Useless Too

The German Bundeskriminalamt (Federal Crime Agency, BKA) recently decided to try out one of those face recognition programs on CCTV cameras placed in a railway station in the city of Mainz. And what happened? Well, having installed the software in October last year, they recruited 200 regular travellers as volunteers, whose faces were recorded in the database.

And the three systems tested successfully identified them 30 per cent of the time on average, with a best result of 60 per cent. Wonderfully, the results were best in the middle of the day - to put it another way, when there were fewest travellers. In the morning or evening, exactly when most of the 23,000 passengers on an average day were on the move, the results were as low as 10 per cent. The hit rate on moving targets was 5-15 per cent lower across the board.

Says the head of the BKA, Jörg Ziercke, "I won't reach the goal of preventing anything with such a low hit rate." He gave a figure of "near 100 per cent" as a minimum, and said he would advise the Minister of the Interior against such systems. The installation in Mainz has apparently been shut down.

The effectiveness of these and similar techniques is something of a bitter question. Since the CCTV boom of the 90s, several British jurisdictions have experimented with recognition. Famously, the system in the London Borough of Newham was exposed as never having caught anyone, and failed a challenge to spot a Guardian reporter even though his presence was announced in advance. Heathrow Airport also employed a system. Results are difficult to come by.

But the Home Office's closed-door trial of various biometric identification methods does give us some data. Their face-recognition software apparently failed in 30 per cent of cases; terrible enough, given the numbers of people the national ID card scheme is meant to process. But it seems to have been dramatically and suspiciously better than the German one - probably just an artefact of its being done under lab conditions rather than in the wild.

It can't possibly work, because biometrics don't scale.

1 comment:

Surreptitious Evil said...

"because biometrics don't scale"

Indeed, at least not as an identification technology - the "birthday paradox" would come rapidly into effect even if the technology functioned properly.

However, as an validation technology (i.e. yes, you appear to be who you claim to be), especially linked with smartcards containing (some of) the relevant ID & biometric data (as opposed to the preferred enormous database) - they do function. Whether they would prove any benefit to society is a separate and altogether more hideous question.


kostenloser Counter