Studies of Operational Measurement of ROC Curve on Large Fingerprint Data Sets Using Two-Sample Bootstrap Studies of Operational Measurement of ROC Curve on Large Fingerprint Data Sets Using Two-Sample Bootstrap

From the operational perspective, on large fingerprint data sets, a receiver operating characteristic (ROC) curve is usually measured by the true accept rate (TAR) of the genuine scores given a specified false accept rate (FAR) of the impostor scores. The ties of genuine and/or impostor scores at a threshold can often occur on large fingerprint data sets, and how to determine the TAR at an operational FAR is provided. The accuracy of the measurement of TAR at a specified FAR for an ROC curve is explored using the nonparametric two-sample bootstrap. The variability of the estimates of standard error and lower bound and upper bound of 95% confidence interval of two-sample bootstrap distribution of the statistic TARs on large fingerprint data sets is extensively studied empirically. Thereafter, the number of two-sample bootstrap replications is determined. Both high-accuracy and low-accuracy fingerprint-image matching algorithms are taken as examples.