[sacw] New Scientist Feature: Caught on camera

[Harsh Kapoor]

=46YI
// Harsh Kapoor//
---------------------------

The New Scientist

http://www.newscientist.com/ns/19990925/caughtonca.html

Caught on camera

You can run, but you can't hide. Big Brother can recognise your face even
if you're in disguise, says Nick Schoon

EVERY DAY, EVERY MINUTE, video cameras scan the
crowds in a busy shopping centre. But this is no ordinary public
surveillance system. The faces of all those passers-by are being converted
into digital code and processed by computer. Continuously, instantaneously,
the facial code of each stranger is compared with that of several dozen
local criminals. If there is a match, an alarm sounds in a control room, a
human operator double-checks the computer's assessment and alerts the
police. Someone they want to keep tabs on, or someone they want to arrest,
is in town...

This is not a scene from some Orwellian future. The system has been
running for a year on the streets of Newham in East London--the first trial
of its kind in the world. And the potential this kind of system doesn't
stop there. Last month, several large stores in Regent Street, London,
began using a system that includes a database of convicted shoplifters who
operate in the area. When a store detective catches someone they think has
been stealing and takes them to the manager's office, a camera mounted on
the wall captures their face, which is automatically checked against the
database. You would also expect this technology to be leapt on by state
security services, such as those unseen watchers who scan airport crowds
for terrorists and drug smugglers. And sure enough, a face-recognition
system devised by Neurodynamics, a company based in Cambridge, England, is
being tested in secret at a British airport.

Facial recognition technology creates new opportunities for companies and
government agencies that want to keep tabs on people. But it is an infant
technology which some fear may be trying to run before it can walk. Is it
reliable enough to be used in such a sensitive field as public safety? And
have our rights to privacy been adequately debated?

We are our faces. To our fellow human beings, if not to ourselves, they
are the key identifiers. Our brains have exquisite machinery for processing
and storing a particular arrangement of eyes, nose and mouth and for
picking it out from other very similar arrangements. This ability is now
being passed on to computers.

True, facial recognition systems have worked for years under ideal
conditions in labs. But they are now moving out. Recent increases in
processing power and improved algorithms can give even fairly ordinary PCs
the ability to capture faces in the hustle and bustle of the real world.
They can spot people who are on the move and are not facing square on to
the camera, and they can compensate for changing light levels. Visionics of
New Jersey, the company behind the trials in Newham and Regent Street,
claims that its technology is not even fooled by hats, spectacles, beards
and moustaches (see "<#1>I see you").

We know who you are

Surveillance is not the only way this technology can be used. In fact much
of the impetus behind it comes from a different branch of the security
industry--one which wants ways to make sure we are who we say we are.
"Biometric" features such as faces, fingerprints, irises, hand geometry and
DNA are powerful identifiers, and facial recognition offers important
advantages over its rivals. It is remote, quick and convenient--the machine
equivalent of a cool glance from a stranger. Many people don't like having
to touch things that thousands of others have fingered and they are
certainly not keen on parting with bodily fluids.

So, when employees in Texas withdraw their pay cheques at hole-in-the-wall
cash machines, a camera and computer make sure that their faces and PIN
codes match. On the Mexican border, people can "fast track" their way into
the US after submitting their faces to automated scans. And in Georgia,
digital images of applicants for a driving licence are checked against the
facial codes of the millions of other licence-holders. (The system has
already detected the same face on more than a dozen different licences.)
But what is so special, and slightly sinister about facial recognition
technology is that people need never know that their identity is being
checked, which is where many surveillance projects begin.

The project in Newham is slightly different in that the local council and
London's Metropolitan Police want criminals to know they're being watched.
The system works by picking out as many faces as possible from passing
crowds, demonstrating that it has "acquired" a face by placing a red ring
around it on a monitor in Newham's closed circuit television control room.
If the software finds a good match between that face and one in its
database of mug shots, the ring turns green and an alarm sounds.

On another screen, a close-up of the face on the street flashes up
alongside that of the criminal. Then human judgment comes into play. Only
if Newham's civilian operator in the control room considers that the two
faces are identical does she or he phone the local police station. For the
police, the system is a potentially useful tool for gathering intelligence
about local villains as much as a way to locate wanted criminals.

Before the Visionics system, called FaceIt, went live last October, Newham
carried out tests to see if it could detect staff members whose images had
been placed on the database. They paraded before the cameras wearing hats,
glasses and other disguises, but the software still recognised them. Since
the system went live, however, it has succeeded in identifying only two
criminals on its database walking the streets of Newham. In both instances,
the police decided not to make an arrest.

There are two possible explanations for this low detection rate. The first
is that villains were so alarmed when they heard about this new
crime-fighting technology that they decided to stay away and have done so
ever since. The second is that criminals are still coming into the area but
the system isn't spotting them. Bob Lack, Newham's head of security, is
hedging his bets on which explanation is correct. He is delighted that
crime in the target area has fallen and believes it could be acting as a
deterrent. But he also accepts that the software still needs improving.
Criminals do, it seems, have a good chance of going undetected, although
not surprisingly Lack and the Metropolitan Police are reluctant to discuss
this.

FaceIt was initially attached to only six of the 154 video cameras
scanning Newham's streets, and it does not acquire every face "seen" by
those six. So what was originally intended as a six-month trial has been
extended, apparently indefinitely. The aim is to have it linked to more
cameras and adapted for "multi-heading"-- acquiring three faces at a time
from the passing crowds. Newham is seeking Home Office funding for this
expansion; in the meantime, the system remains operational.

Lack is keen to stress the multitude of measures designed to prevent the
system being abused. The control room is closely guarded and the digitised
faces of men and women in the streets are discarded as soon as the system
finds no match. Furthermore, the control room operators see only the faces
of criminals on the database: they do not see their names or records. That
information stays with the police, who will not say who is on the database
or even how many offenders are included.

Nothing to fear?

An internal police committee chooses the list of names, which probably
embraces paedophiles and drug dealers as well as violent robbers and
burglars. David Holdstock, a spokesman for the Metropolitan Police, says
people convicted of trivial offences, or who were not regarded as active
criminals, would never go on the database. "It's pretty focused and pretty
targeted," he said. "If you're innocent you have nothing to fear."

Others are not convinced by this reassurance. When the project began, the
Metropolitan Police reckoned that the system made accurate matches in 80
per cent of cases, says Liz Parratt of Liberty, Britain's council for civil
liberties. "I see that as a 20 per cent failure rate," she says. "That's an
unacceptable level of error." How many innocent people must have their
collar felt for every criminal caught?

There's also the more general issue of whether we are watched enough
already. Britain has become the surveillance capital of Europe, with
cameras in virtually every city and large town, and many smaller ones.
"Everyone wants to see crime fall," says Parratt, "But I'm not sure I want
to live in a society in which when I walk down the road my face is compared
with a couple of hundred local criminals."

Robin Wales, the elected leader of Newham Council, sees things the other
way round. "There is a civil liberties issue at stake here," he says. "It's
the right of our citizens to live in Newham without fear of crime." He
claims widespread local support for its decision to introduce the
technology.

Britain's Data Protection Registrar, whose task is to protect the interest
of individuals against data collectors, has scrutinised Newham's system and
is satisfied with the safeguards. But what happens next might be a problem.
If the trial is eventually pronounced a success in reducing crime then it
may well be extended to other public surveillance systems, first in Britain
and then elsewhere. And if these systems share a large, common database
then criminals on it would, in effect, be "electronically tagged". The
police would know whenever they showed their faces in any place covered by
cameras.

Clearly this could be a powerful deterrent to crime, but it would take us
a big step nearer to a world in which Big Brother really is continuously
watching everyone. "We would have serious concerns about such a general
system," says Jonathan Bamford at the Office of the Data Protection
Registrar. "We think it would be inappropriate." After all, the decision to
sentence offenders to be electronically tagged, which gives the authorities
a constant check on their location, is made by the court, not the police.

So the spread of this technology raises an old but important question: who
guards the guardians? The actions of police, other state authorities and
private companies in compiling and maintaining face databases needs to be
kept under scrutiny. How serious a crime do you have to commit before your
face goes on a system? How long do you have to stay clear of crime before
your face is removed? Do you have the right to know that you are on a
database? In Newham, there has been no debate about these issues.

These questions are not just of interest to the criminally-inclined. A
small but fast growing number of PCs are sold with an integral video camera
facing the operator. One potential use for this is to check a person's
identity when they log on to a network or try to buy something over the
Internet. Chris Solomon, who works on facial recognition at the University
of Kent in Canterbury, believes this kind of security system will be
commonplace within a few years.

Visionics' rival Miros, a company based in Wellesley, Massachusetts, is
already marketing software that gives access to secure Web servers only to
those who have registered their faces. Solomon sees great advantages and a
host of applications in the technology. "But there are concerns," he says.
"Like so many things in life, it's a stick with two ends."

I see you

BOTH IN GREY MATTER
and in silico, facial recognition is a two-stage process. First you need to
detect that a face is in the visual field. Then you need to lock onto it
and rapidly compare it with a database of thousands of faces in memory.
This process needs to allow for the fact that faces differ from moment to
moment and year to year. The most obvious source of variation is the angle
from which it is seen, but there are others, from changing light levels and
skin tones to expressions and facial hair.

Our brains overcome these confusions, allowing us to recognise a face with
astonishing rapidity, along with a host of accompanying
information--favours owed or owing, character assessment and (if we're
lucky) a name. Algorithms capable of mimicking this astonishing ability, at
least in part, improved through the 1990s. Increases in computing power and
the reduction in cost of computers and video technology is starting to make
facial recognition affordable.

Several different approaches to automatic face recognition have emerged.
One builds up a person's face by overlaying many different facial types,
known as eigenfaces. A computer varies the brightness of every eigenface in
turn until it finds a combination that resembles the face. The set of
numbers representing the brightnesses of the eigenfaces can then be
compared with other sets stored in memory (Sleuth City, New Scientist
supplement, 4 October 1997, p1).

The Visionics system, deployed in Newham, works by dividing up the image of
a face into between 12 and 40 elements, such as the outer corners of the
eyes, tip of the nose and ends of the eyebrows. It then measures the
distances between these elements, and gauges the angles between lines that
are drawn between them. These numbers become the digital identity of a
face. The software can compensate for head movements and changes in
expression because it "knows" that the elements can move slightly relative
to one another.

The system focuses on a triangle between the base of the nose and the outer
edges of the eyebrows so, say the company, moustaches and beards do not
confuse it. Nor, it claims, do spectacles--unless heavy shades or glare
occludes the eyes. Faces can be turned up to 15 degrees away from the
camera without any loss of recognition performance. Once that angle is
exceeded, its ability to identify starts to deteriorate.

Of course, before the work of recognising can begin, the system has to find
a face within the video camera's field of view. It does this with another
set of pattern-matching algorithms which detect "head-shaped" objects in
the field of view. Then, once a face has been detected, it is "normalised"
to compensate for variations in lighting condition, size (or distance from
the camera) and skin colour. This is rather like taking a photograph of a
crowd, enlarging each face in it to fill the same size of frame and
adjusting the contrast and the average shade of each to standard values.

What can you do to fool a face recognition system? Sporting a bag over your
head might arouse suspicion. Wearing heavy shades and a daft expression
would almost certainly throw the software off the scent, but you'd just
have to hope that you didn't bump into someone you knew.

>From New Scientist, 25 September 1999

=A9 Copyright New Scientist, RBI Limited 1999