Robert Važan

Numbers in fingerprint recognition systems

Beginner introduction to common fingerprint recognition terms like 1:1 and 1:N, FAR, FRR, and EER.

Fingerprint recognition is all about numbers. Aside from some minor features, matching algorithm is predominantly defined by its accuracy and speed. Numbers are attractive, because they are concrete and they are easy to compare. There are caveats though. If you happen to compare meaningless numbers, you will get meaningless result.

I've already written why speed doesn't matter beyond some basic speed level. Chiefly, the reason is that hardware is indeed dirt cheap if you know how to rent it efficiently.

Only one 1:N identification per user is really needed in order to deduplicate the user. All the other queries are positive identifications where candidate set is small, often just one candidate fingerprint.

FAR can be driven arbitrarily low by increasing matching threshold. With quality fingerprints in the database, it is possible to set threshold high enough to make it nearly impossible to get a false match.

FRR is irrelevant for just about every real-world application. Deduplication doesn't need low FRR, because even 90% probability of being caught is enough to discourage people from cheating. Quality fingerprints have essentially zero FRR. Most of the FRR comes from poor sensors or poor scans, both of which can be eliminated in positive identification applications.

EER reached in competitions is irrelevant, because everyone has thresholds set for at least FAR10000 and usually much higher. Most algorithms have fairly flat FRR within sane range of thresholds.

Competitions in general do not matter much, because companies submit ultra-slow algorithms there that maximize accuracy. Then they sell you the fast version with much worse accuracy. But then competitions are better than marketing bullshit on vendor websites. They can at least show vendor's ultimate limits.

Accuracy on a big database is a whole different matter than accuracy of simple 1:1 matching. Additional verification is required at the end of the search to prevent error accumulation, but even then extrapolating 1:1 to 1:N accuracy is nontrivial.

That's why SourceAFIS accuracy is good enough for so many users. Sure it could be better, but it's not essential. It's just for comfort as extra accuracy and extra speed allow users to use the technology in wasteful ways to save their own time.

It's mostly the government contractors who push big numbers, because they expect big bucks in return. Governments have lots of money they are unable to spend efficiently. Lots of times they just buy excess luxury to make their lives easier.