Abstract: Biometric identification uses personal traits and characteristics such as fingerprints and facial recognition to identify an individual. As technology becomes increasingly digital and security concerns gain in prevalence, this field will continue to grow rapidly. Biometric identification devices are not foolproof; they produce both false acceptances and false rejections. Our study aims to improve the inferential estimation of the false accept rate and the false reject rate. In order to do this we assume the proportions in practice are generated according to a Beta-Binomial distribution. Our aim is to create the optimal parametric confidence intervals for biometric error rates. In particular, we are interested in situations with low error rates (0.001 and lower). Using transformations such as the logit, arcsin of the square root, and Wilson transformations, we evaluate the performance of these confidence interval approaches. We find that the logit and the arcsin of the square root transformations are very similar and all three are better than the traditional Wald confidence interval. Additionally we examine combining the transformations and discover that the best possible transformation is the logit transformation on the Wilson transformation base. |
