[fprint] Fingerprint Misidentification

[simon at koala.ie]

On Tue, 05 Jun 2012 21:45:26 +0100, Keaton Mowery <kmowery at gmail.com>  
wrote:

> Hi all,
>
> I recently wrote a small fprintd using libfprint and deployed it on
> our biometric soda machine at UCSD. Registration works great, and
> recognition of registered fingers is also working fine. Unfortunately,
> we're having some trouble with non-registered fingerprints. Namely,
> they are recognized as other, registered fingerprints.

i can talk a bit more about this now that i no longer work for a  
biometrics company.

what you need to understand is that fingerprints are a ridge pattern on  
top of a rather flexible base (your finger).
this means (coupled with cuts and abrasions) that no two captured prints  
are identical.
matching works by creating a template (some use pattern matching) which  
plots the coordinates of the minutiae in the print.
these are the corrdinates and directions of the start, end and branch  
points of the ridges.
depending on the type of sensor and the environment (heat, light, dust)  
you get a varying number of points.
matching then compares these lists and applies hueristics to determine if  
they are a match.
what is returned is a probability. not a yes/no answer. that can never  
occur.

there are two error rates at play here.

False Accept Rate (FAR)
False Reject Rate (FRR)

what you are having problems with is a FAR which is too high for your  
application.
however, if you increase the threshold you will increase the FRR.

that is the nature of biometrics.
the trick is balancing FAR against FRR.

one thing you can do is match against two prints; say left index and the  
right index.
this will change the operating point. but again you have to balance the  
FAR against the FRR.

hope this is of some use.
-- 
simon

Simon Kenyon
e: simon at koala.ie
m: +353 86 240 0005