Calibrated Confidence Scoring for Biometric Identification
Dmitry O. Gorodnichy, Richard Hoshino · 2010
Existing biometric identification systems, such as those used in trusted traveler programs, attempt to identify an individual’s identity from an enrollment database of n people. The output is either the name of an enrolled person, or a rejection message indicating that no match was found. Traditionally, no measure of confidence is given to the output; an individual is either granted or denied access. In this paper, we propose an extension to existing biometric systems by applying a calibration function to the n matching scores. We introduce a computationally-light calculation that can be applied either as a post-processing filter or embedded directly into an algorithm to yield perfectly calibrated probability-based scores. In addition to attaching a meaningful confidence measure to the output, the proposed methodology is also shown to improve the overall performance of a biometric system. The theoretical proof of the calibration formula is followed by its application to iris biometrics, on a data set consisting of nearly 60, 000 iris images. By comparing the detection error trade-off (DET) curves, we show that our score calibration post-processing filter reduces the area under the DET curve from 2.41 to 0.17, and reduces the equal error rate (EER) from 5.40 % to 2.84%. 1