Quantifying the impact of missing minutiae on latent fingerprint identification

Abstract

Fingerprints are the patterns left behind by the ridges of the skin on the tips of the fingers and are commonly used as an identifying characteristic of individuals. Fingerprints are identified by comparing specific features between them. Minutiae are the most commonly used features for fingerprint identification. In an attempt to simulate and study human error when manually marking minutiae, we performed a study to determine the performance of a fingerprint matcher when it is input a subset of the apparently available minutiae in a latent fingerprint. This situation is analogous to an expert overlooking existing minutiae or misleadingly annotating a minutia when there is none. This was done by removing minutiae that were manually marked by latent fingerprint experts, and comparing matching score and rank-$n$ identification performance to the original expertly-marked fingerprint. We found that randomly removing minutiae from latent fingerprints generally causes the recognition rate to go down in a closed set comparison experiment. We removed all possible combinations of one, two, and three minutiae from latent fingerprints and found that when removing minutiae it is more likely for the matching score to go down instead of up. We also found that in some cases, removing even one minutia can cause a fingerprint not to be identified. And that if removing a set of minutiae from a fingerprint caused a drop in matching score, it was more likely that it would cause a drop in rank-$n$ identification performance rather than an increase. Finally, we created a dataset based on the minutiae removed and their change in scores to train and evaluate several machine learning models to predict how a minutia will affect the matching score of a fingerprint if it is marked. Our best model was able to predict with a better-than-random chance if a minutia will increase or decrease the matching score of the fingerprint with a .601 AUC.