On Controlling Genuine Reject Rate in Multi-stage Biometric Verification

An important problem in multi-stage biometric verification is to select an appropriate reject region. A reject region says which samples to be rejected. Rejecting impostor samples does not incur any cost in terms of user inconvenience, however, erroneously rejecting genuine samples leads to both user and administrator inconvenience. The problem becomes severe in the applications that involve a huge number of biometric transactions. Such applications necessitate the reject rate of genuine samples to be controlled. However, to date, no work has studied on controlling genuine reject rate (GRR) in multi-stage biometric verification. In this paper, we focused on controlling GRR and to this end, developed a rejection method called symmetric rejection method. Our rejection method adds the following benefits to multi-stage biometric verification: (1) it enables the system administrator to control GRR, (2) it allows to calculate the reject region without estimation of score distributions, and (3) it does not use any assumption on the functional form of score distributions. We performed experiments on (1) two fingerprint datasets of 6000 users and (2) two face datasets of 3000 users. For fingerprint data, we achieved 18.96 percent to 70.89 percent reduction in EER by rejecting 1.5 percent to 9.4 percent genuine scores and for face data, we achieved 3.27 percent to 85.83 percent reduction in EER by rejecting 0.3 percent to 14.4 percent genuine scores.