« Uppity women | Main | Disclosure »

Ground truthiness

New term of the week: ground truthiness (or it may be a word, groundtruthiness; I don't hear spaces well). It refers to what happens to a shiny, new technology when it moves from abstract theory and neat, clean lab demonstration into the messiness of the real world, where some people have two right thumbs or pendular nystagmus. In the wild, you find reality; you find the truth of what happens on the ground. You find its ground truthiness.

The ground truthiness of biometrics is particularly fragile. It's not just users that can confound things - as James L. Wayman said in 2003 ($), "People never have what you think they're going to have where you think they're going to have it." Bigger issues are the environment, the context of the system, and the needs and expectations of its owners.

The first kind of story is sometimes laughable: a facial recognition system built and marketed in Nigeria that does not work particularly well for dark faces. The second kind is why when you use the airport electronic gates you have to stand correctly on the yellow feet and look ahead so the lighting is right. The last two are the really complex ones.

On Tuesday, at the 2014 biometrics conference UCL researcher Itiel Dror explained that despite their neutral, objective sound, biometrics systems can be as unreliable as human memory. The work Dror described studying cognitive bias in forensics is the fingerprint analogue of the work done by Elizabeth Loftus demonstrating the vagaries of eyewitness testimony ($) and the bias that can be conveyed in police line-ups. After decades of vendor marketing, we tend to think of biometrics as clean, neutral, unimpeachable, and objective: if a fingerprint lying around the scene of a crime matches yours you'd better have a really good explanation. And yet...

The fact is that however neutral the scrap of data that represents a biometric may be, the system it's built into is designed and operated by humans. In one of the applications Dror studied, matching latent fingerprints collected at crime scenes with fingerprints stored in an automated fingerprint identification system (AFIS), the system gives the examiner a list of possible matches, often 15, sometimes as many as 50. Just as Loftus found that eyewitnesses may be influenced by the police's ordering of a lineup, Dror discovered that the expert examining fingerprint matches is influenced by the ordering of the list and also by the context of what they've been told about the case - for example, knowing that the police considers a particular suspect guilty.

Experts in any field, Dror said, learn what elements they can safely ignore. If you know that most of the time the match is in the top few, you may well miss better matches appearing lower down. There are solutions to this, which he proposed. Randomizing the order is the most obvious (and, for similar reasons, is done in some jurisdictions with candidates' names on election ballots). Dror also suggested giving a prize for matches found lower down the list. However, both strategies slow the examiner's work. His compromise is to suggest giving shorter lists (the length to be determined by studying historical data and setting a threshold based on the number of desired hits) and randomizing the items.

"The success of biometrics systems depends on how well they are integrated into a larger cognitive context," he concluded.

In his paper, Dror suggests that this is a new problem that has been introduced by the creation of the giant AFIS databases: in earlier times there tended to be only a few suspects and those were selected based on other criteria than the probability of a biometric match. A similarly new problem, though different in kind, was introduced in India, said Nishant Shah, cofounder of the Centre for Internet and Society, who has been working with the Unique Identification Authority of India. Part of UDAI's goal is to help severely disenfranchised groups such as the homeless, nomadic workers, and asylum seekers. But, he said, "Some communities only survive because they are not counted." And there are other difficulties: village manual labourers do not have legible fingerprints, while in one southern village only the upper castes were registered. The reason: the enrolment equipment was set up on land owned by upper castes, to which the lower castes had no access.

Shortly afterwards, Alan Gelb, from the Center for Global Development, noted that the introduction of systems that tighten the link between identity and citizenship is increasing the number of formally stateless persons, currently estimated at 12 million worldwide. "The privacy gap has gotten a lot of attention," he said, "but there is very little on the risk of exclusion."

None of these problems is specifically about biometrics. But they are all about the difficulties inherent in trying to fit the fuzzy arrangements that humans make among themselves into tidy binary categories. It is a much more profound change than most vendors recognize - especially those like Jim Jasinski, senior vice-president for business development at Digital Signal, who repeatedly insisted that the technology his company sells is "just a conduit".

Wendy M. Grossman is the 2013 winner of the Enigma Award. Her Web site has an extensive archive of her books, articles, and music, and an archive of earlier columns in this series. Stories about the border wars between cyberspace and real life are posted occasionally during the week at the net.wars Pinboard - or follow on Twitter.


TrackBack URL for this entry:

Post a comment

(If you haven't left a comment here before, you may need to be approved by the site owner before your comment will appear. Until then, it won't appear on the entry. Thanks for waiting.)