I CAME here looking for an argument but I can鈥檛 find one. All round this lofty exhibition hall 鈥� billed as the world鈥檚 biggest market for security equipment 鈥� the people selling face-recognition systems are being disarmingly, infuriatingly honest. OK, they don鈥檛 make the systems themselves, they just work for security equipment firms and are selling the products on. But I thought they鈥檇 at least attempt to defend the technology. When they don鈥檛, it鈥檚 me who鈥檚 caught off guard.
Is it true that the systems can鈥檛 recognise someone wearing sunglasses? Yes, they say. Is it true that if you turn your head and look to one side of the camera, it can鈥檛 pick you out? Again, yes. What about if you simply don鈥檛 keep your head still? They nod.
Maybe nine or ten months ago they would have risen to the bait. In those days the face-recognition industry was on a high. In the wake of 11 September, Visionics, a leading manufacturer, issued a fact sheet explaining how its technology could enhance airport security. They called it 鈥淧rotecting civilization from the faces of terror鈥�. The company鈥檚 share price skyrocketed, as did the stocks of other face-recognition companies, and airports across the globe began installing the software and running trials.
Advertisement
As the results start to come in, however, the gloss is wearing off. No matter what you might have heard about face-recognition software, Big Brother it ain鈥檛.
The first warning came from within the industry. In January, an article in Security Management magazine cautioned that the industry would live to regret the hype. It was written by researchers at Image Metrics, a British company that develops image-recognition software. They warned of the danger of exaggerated claims, saying that 鈥渁n ineffective or poorly applied security technology is as dangerous as a poorly tested or inappropriately prescribed drug.鈥�
Their concern was based largely on an independent assessment of face-recognition systems carried out in 2000 in the US by the Department of Defense. These tests found that to catch 90 per cent of suspects at an airport, face-recognition software would have to raise a huge number of false alarms. One in three people would end up being dragged out of the line 鈥� and that鈥檚 assuming everyone looks straight at the camera and makes no effort to disguise themselves.
Results from the recent airport trials would seem to justify that concern. One of the first reports came from Fresno Yosemite International Airport in California, which was trialling a system made by Viisage of Littleton, Massachusetts. A bearded man of Middle Eastern appearance spent several hours being interrogated by the FBI after the system flagged him up as a terrorist suspect. By the time the unfortunate traveller managed to convince the Feds he was not a threat, he had missed the last plane.
Then in May, Palm Beach International Airport in Florida released the initial results of a trial using a Visionics face-recognition system. The airport authorities loaded the system with photographs of 250 people, 15 of whom were airport employees. The idea was that the system would recognise these employees every time they passed in front of a camera. But, the airport authorities admitted, the system only recognised the volunteers 47 per cent of the time while raising two or three false alarms per hour.
In July came another blow. A security consultant hired to test face-recognition systems for Boston鈥檚 Logan Airport reported that the security staff operating the systems were being overwhelmed by false alarms.
So why all the problems? How come computers find it so hard to match a picture of a face with its owner?
Most face-recognition systems use some kind of geometric technique to translate a picture of a face into a set of numbers that capture its characteristics. Different companies have slightly different methods, but the principle is broadly similar (see Graphic). The software takes a digitised image and scans the pixels, looking for areas of high contrast. These usually indicate some kind of boundary: an eye socket, a cheekbone, or the lips, nose or hairline, for example.
Once it has identified these boundaries, the software calculates their relative sizes and positions and converts this geometry into what Visionics calls a 鈥渇aceprint鈥�. Feed the software a series of mugshots 鈥� from a terrorist watch list, say 鈥� and it鈥檒l calculate their faceprints too. Then it can monitor live CCTV images for the faces of known suspects. When it finds a match, it raises an alarm.
That鈥檚 the theory, but in the real world it鈥檚 not quite so simple. The ambient light level and the quality of image from the camera are key factors. They will determine how many points of contrast the software can pick out from the image. The angle at which the face is presented to the camera also matters: if the software can鈥檛 find two eyes it will struggle to capture a faceprint at all. The number of images in the database also has an effect. Large numbers slow down the search for matches, or produce too many possible matches.
The airport trials show how these simple problems can throw a face-recognition trial out of kilter. At Palm Beach, for example, anyone moving their head caused the system problems, and if a face was morethan 15 degrees off 鈥渇ace on鈥� to the camera in any direction, there was a substantial loss in matching. Glasses of any sort caused problems, with glare making capture difficult.
Even if the system does manage to capture a face, the problems aren鈥檛 over. Checking for matches isn鈥檛 straightforward, either. The trouble is that a suspect鈥檚 faceprint taken from live CCTV is unlikely to match the one in the database in every detail. So somewhere along the line there are choices to be made over how close the match has to be to trigger an alert. Do the two faceprints have to match almost exactly? Or will a vague resemblance do?
As with any security system, the trick is to navigate between two undesirable extremes. On one side are too many false positives 鈥� innocent passengers being hiked out of line by the security staff. On the other are false negatives that allow known terrorists to swan through unchallenged because the system didn鈥檛 find an exact match.
False alarms
To give themselves the best chance of picking up suspects, operators can set the software so that it doesn鈥檛 have to make an exact match before it raises the alarm. But there鈥檚 a price to pay: the more potential suspects you pick up, the more false alarms you get. You have to get the balance just right.
Visionics 鈥� now called Identix after merging with a fingerprint-scanning company in June 鈥� is quick to blame its system鈥檚 lacklustre performance on operators getting these settings wrong. Sure, Palm Beach was something of a PR disaster, says Identix spokeswoman Frances Zelazny, but that鈥檚 only because the airport authorities fixed the settings to minimise false alarms. Anyone who does that is bound to let lots of matches slip through the net. She reckons that to get optimal results you have to accept a false alarm rate somewhere between 1 and 3 per cent. 鈥淚n Palm Beach it was almost zero: that鈥檚 not acceptable.鈥� The company prefers to cite another trial, at Dallas/Fort Worth Airport, which used less stringent settings. There, with a 1.2 per cent false alarm rate, the system correctly identified faces on its database 94 per cent of the time.
Viisage uses a similar argument to claim the Logan Airport trials were actually a success. The software is designed to alert security guards to possible suspects so they can make visual checks, says chief marketing officer Cameron Queeno. Guards can鈥檛 keep the details of 2000 faces in their heads, but they can compare a passenger against mugshots displayed on a screen. If that seems labour intensive, it鈥檚 no different from what happens at the baggage X-ray machine. The software is just an aid, not a solution, Queeno says.
But it turns out that even this creates problems. The consultant responsible for the trial at Logan reported that the majority of passengers and airport workers prompted the system into suggesting a possible match. This burdened the security guards with the task of scrutinising thousands of faces against endless photos of similar-looking people, and in the end they were overwhelmed.
This shouldn鈥檛 come as a complete surprise. Numerous studies have shown that people are surprisingly bad at matching photos to real faces. A 1997 experiment to investigate the value of photo IDs on credit cards concluded that cashiers were unable to tell whether or not photographs matched the faces of the people holding them. The test, published in Applied Cognitive Psychology (vol 11, p 211), found that around 66 per cent of cashiers wrongly rejected a transaction and more than 50 per cent accepted a transaction they should have turned down. The report concluded that people鈥檚 ability to match faces to photographs was so poor that introducing photo IDs on credit cards could actually increase fraud.
The way people change as they age could also be a problem. A study by the US National Institute of Standards and Technology investigated what happens when a face-recognition system tries to match up two sets of mugshots taken 18 months apart. It failed dismally, with a success rate of only 57 per cent.
There鈥檚 another fundamental problem with using face-recognition software to spot terrorists: good pictures of suspects are hard to come by. The detainee at Fresno was matched against a database of mugshots taken from the TV programme America鈥檚 Most Wanted. It seems laughable, but that was the best source available, as the main photographic terrorist databases in the US are held by the CIA. Some of those images can鈥檛 be released into the public domain because they are of innocent people: the security services are concerned about them, but don鈥檛 have the evidence to make an arrest. Very few security personnel at American airports have CIA clearance, so they aren鈥檛 allowed to see the images. 鈥淯ntil they鈥檝e got cleared personnel in each of those airports they can鈥檛 stop terrorists getting on planes,鈥� says Iain Drummond, chief executive of Imagis technologies, a biometrics company based in Vancouver, Canada.
Despite the disappointing tests, Queeno insists that face-recognition technology is good enough to put terrorists off. 鈥淎re they going to walk into an airport that has the technology deployed and say, 鈥榃ell, there鈥檚 only a 90 or 80 or 50 or 40 per cent chance I鈥檒l be stopped鈥�? Absolutely not. They鈥檙e going to go someplace where there鈥檚 a zero per cent chance they鈥檙e going to get caught. The technology is helping.鈥�
Airport security isn鈥檛 the only use for face-recognition software: it has been put through its paces in other settings, too. One example is 鈥渇ace in the crowd鈥� on-street surveillance, made notorious by a trial in the London Borough of Newham. Since 1998, some of the borough鈥檚 CCTV cameras have been feeding images to a face-recognition system supplied by Visionics, and Newham has been cited by the company as a success and a vision of the future of policing. But in June this year, the police admitted to The Guardian newspaper that the Newham system had never even matched the face of a person on the street to a photo in its database of known offenders, let alone led to an arrest.
Identix says the system works as a deterrent, and that鈥檚 all the company ever claimed. But some people in the industry say face recognition isn鈥檛 up to being used on the street. 鈥淔ace recognition works best in a controlled environment,鈥� says Drummond. 鈥淲hat it doesn鈥檛 do is face-in-the-crowd stuff 鈥� that鈥檚 quite a long way off. We don鈥檛 encourage our users to do face-in-the-crowd at all.鈥�
Camera shy
There have been other, none-too-successful trials of street surveillance. Earlier this year, Britain鈥檚 Essex police tried out a Visionics system against shoplifters. It did at least make a couple of matches with a database of known offenders. But in the end the police abandoned the system. Face recognition was certainly impressive in theory and even in controlled trials, they said. But it just wasn鈥檛 ready for the real world.
The problem in Essex was with the CCTV cameras in the shops. Mounted near the ceiling and looking over a wide angle from a considerable distance, they鈥檙e no good at spotting faces. Put a baseball cap on firmly, and no camera will catch your mug. Add to that the fact that the cameras are usually several years old and recording on tapes that have been used hundreds of times before, and the image quality is not good enough.
Given the right conditions and the right application, however, face recognition can be useful. Many police forces are using it when booking in suspects they have arrested. With tightly controlled lighting, and detainees compelled to look straight into the camera, face recognition has succeeded in linking suspects who gave false names with mugshots held in police databases up and down the US. The National Crime Squad in Britain and customs authorities in New Zealand have bought Imagis face-recognition systems. Despite its problems and embarrassments, face recognition seems to be here to stay.
But the feeling persists that Visionics鈥檚 claims after 11 September were premature. 鈥淲e have to bring back expectations to the right level,鈥� Drummond says. Someone may just be about to do that. The US Department of Defense has finished a new assessment of commercial face-recognition systems and plans to release the results later this year. After all the claims and counter-claims, with no one able to discern the truth, the industry may soon have to face up to reality.