Studies of Biometric Fusion NISTIR 7346; Appendix B: Effectiveness of Score-Level Fusion

This three-part appendix contains the results of experiments measuring the effectiveness of different categories of fusion: multi-modal (finger and face), multi-instance (multiple finger positions), multi-matcher, and multi-sample (multiple enrollments). Appendix B.1: Score-Level Fusion of Face and Multiple Fingerprints This is an analysis of the effectiveness of multi-modal (finger and face) and multi-instance (multiple finger positions) score-level fusion, focusing on the extent to which different biometric modalities and instances are independent, and the effect of that independence on the accuracy of fusion. It includes detailed analyses of the effects of fusing scores from varying combinations of fingers, and the effect of fusing face and fingerprint scores. This paper provides large-scale empirical evidence that score-level fusion using multiple finger positions is highly effective, as is fusion of fingers and face: fusing two fingerprints or one fingerprint and face generally resulted in a 50-90% reduction in false reject rate (FRR) relative to the stronger of the two inputs at a constant false accept rate (FAR). Appendix B.2: Score-Level Fusion of Multiple Matchers This is an analysis of the effectiveness of score-level matcher fusion, in which multiple matchers produced scores from comparisons of the same pairs of images. Both face and fingerprint matchers were evaluated. Any improvements in accuracy reflect differences in the matchers that might be exploited either through score-level fusion or further improvement of existing matcher technology. A 10-30% reduction in missed identifications (relative reduction in false rejection rate) was achieved. Due to data correlation, algorithm fusion is less effective than either instance or mode fusion, but can still improve accuracy given limited data. Appendix B.3: Score-Level Fusion of Multiple Fingerprint Samples This is an analysis of the effectiveness of score-level sample fusion, which uses more than one sample from each biometric instance, such as multiple fingerprint images from each of a person’s fingers. Multi-sample fusion is of operational interest because it can improve matching accuracy without additional collection of data by retaining successfully matched probes in the gallery in addition to the originally enrolled sample. False reject rates were reduced by 45% to 73%. Studies of Biometric Fusion