An Effective Two-Finger, Two-Stage Biometric Strategy for the US-VISIT Program

Motivated by the cost and disruption involved in changing from a two-finger to a ten-finger biometric system for matching U.S. visitors to a watchlist of criminals and terrorists, we investigate whether any two-finger multistage biometric strategies would fix the inadequate matching performance of poor-quality prints that plagues the U.S. Government’s original twofinger, single-stage biometric system. For several multistage strategies, we solve the Stackelberg game in which the U.S. Government chooses the biometric threshold levels to maximize the detection probability subject to constraints on the false positive probability and on the mean time per visitor to perform biometric screening, and the terrorist chooses the fingerprint image quality to minimize his detection probability. The first stage of all the strategies uses the current minutiae-based fingerprint matching system, but with thresholds that depend on image quality, which in isolation achieves a detection probability of 0.771. Using face recognition (based on 2002 data) in the second stage increases the detection probability to 0.841, whereas using a slower and more thorough texture-based fingerprint matcher in the second stage leads to a detection probability of at least 0.913 and perhaps significantly higher. (Data for the texture matcher is only available for the poorest-quality prints and we assume that this is its performance for all prints.) Adding face recognition as a third stage to this latter system does not improve performance. The two-finger, two-stage strategy may be comparable in performance to the ten-finger, single-stage strategy (which has a detection probability of 0.937), is robust against gaming and poor image acquisition, requires no additional hardware, and would generate no visible changes from the original two-finger, single-stage system from a visitor’s viewpoint. The uncertainty in our performance estimates needs to be better quantified, ideally with raw data on similarity scores, before recommending this strategy for implementation.