Facial Recognition. Three common pitfalls, and how to fix them. Pitfall #1

We do not understand what quality photos we have in the comparison database, and we cannot adequately build decision thresholds “yes / no”. Thresholds are set either "by eye", or based on some general recommendations.

The result is not optimal operation of the recognition system, a multiple increase in errors of the first and second kind. Additional cost for processing erroneous comparisons.

The cost per mismatch incident is highly dependent on the business case.

If we talk about ACS, then this is the time spent by the security service on re-identification in case the correct person is not missed, or the time and effort spent on localizing the consequences of the penetration of a “stranger”.

If we take a more complex case - identification during a banking transaction, then this is the price of a possible loss of a client if we do not recognize him, or the admission of a stranger to the user's bank accounts.

Let's consider a business case of bank application login through biometrics.

According to statistics, the average number of logins per user is 19 to 25 times a month. Take, for example, a bank with a base of 1 million customers. We receive at least 19 million identifications per month. The best NIST algorithms for the highest quality Visa Photos dataset show FNMR = 0.0006 at FMR = 0.00001

In practice, having a dataset that is not quality checked, we will get at best FNMR = 0.003 with FMR = 0.00001.

This is more than 45,000 "extra" incidents per month. If each incident costs us an average of $1 (which is a grossly underestimate), then we get from 450,000$ per year of "extra" costs**. This amount is quite commensurate with the annual cost of owning a biometric system.

Using the photo quality control algorithm, we build a procedure for regularly checking the dataset. We remove and replace low-quality photos.

Cost savings commensurate with the cost of ownership of a biometric service, or 0.002$ per 1 transaction.