Recently, I checked in with James L. Wayman, one of the longer-server researchers into biometrics to ask on the progress of work I had heard he was doing on multiple biometrics. That is, systems which use more than one biometric to authenticate identity. It seems logical to assume that two biometrics would be more accurate than one. In practice, no, privacy advocates have argued to me, because although having multiple biometrics makes the sense more privacy-invasive, in practice one biometric is bound to be more reliable than the other and people will rely on the one that’s more reliable, so there’s no accuracy benefit. Now that my fingertips are practically smooth, I disagree with this: a system that used both iris scans and fingerprinting would be more accurate in my personal case than one that just used fingerprints - 8 of mine are now so poor that they wouldn’t be accepted for enrollment.

In any event, in response to my query Jim sent a copy of this paper, which is beautifully clearly written, in between the pretty, decorative mathematical formulas and equations. The gist: more actually is better, in terms of improving accuracy. However, despite 30 years of research, no one is using multimodal systems because they are expensive to install; they are complex to manage (it’s hard enough to get the right lighting and camera work to do, say, just facial recognition; it’s considerably more difficult to get the right environment for two different types of biometric that have different demands); and these systems are difficult to test because of the privacy implications of collecting and publishing so much personal detail about volunteers.

What is being used and is successful is multi-presentation (but single mode) biometrics (10 fingers instead of 1 or two, two irises instead of one), or multi-instance (but single mode) biometrics (five cameras capturing the face from different angles instead of one), or even multi-algorithmic (for example, using to different speech recognition engines for speech). These do indeed improve accuracy.

This is true despite the fact that these biometrics may be correlated; in other words, they are not entirely independent of one another. The pattern in your iris is not correlated to the pattern of your fingerprints. But the patterns of your ten fingerprints are correlated with each other; that is, thay have similarities (Wayman quotes Galton on the Bertillon system, pointing out that the various meausrements that make up the system are correlated - a tall man is more likely to have a large foot, etc.).

The biggest improvement in error rates is achieved by improving the quality of the information captured (that is - better quality fingerprints; five images instead of one).

wg