Accuracy rates vary with threshold settings. For applications with higher stake outcomes, for example transferring $100,000 from one bank account to another, would require a higher security threshold. The IT team can adjust settings to set a higher bar for false acceptance rates, so it’s less likely that a user can get through the authentication without a near perfect match on the face recognition. Likewise, for applications where the stakes are lower, for example logging into Netflix, the false rejection rate can be lowered which enables a match on a less perfect image which should enhance the user experience.
As the false acceptance rate (FAR) is raised, the false rejection rate (FRR) is lowered. Raising the FAR to 1 in 1000, FaceLocate‘s FRR’s rate is 4.5%, meaning if we want to ensure that there’s only a 1 in 1000 chance that our system will incorrectly say yes, that’s the account owner, give them access, the trade off is that the system is set to such a high bar for making a match, so about 4.5 times out of 100 it WILL be the correct user but the system will say it’s not.
It’s critical that application developers and the enterprise have the ability to set these thresholds, to achieve the right tradeoffs between security and ease of use, specific to their use cases.
One way to establish a universal accuracy rating is to set the threshold a equal for both false matches and false non-matches. FaceLocate equilibrium balance point between false acceptance and false rejection is 1.3 % for the FERET database. FaceLocate flexible parameter settings allows for higher security or greater ease of use with a 99.99 % security level easily achievable.