c©Copyright by

Although fingerprints have been used in forensic identification for over a century, establishing the degree of uniqueness a given fingerprint has remained elusive since it involves relation-ships and correlations hidden in a large amount of fingerprint data. We develop machine learning approaches to this problem focusing on two aspects: better generative models for fingerprints, and estimating probabilistic metrics for individuality and rarity. Generative approaches are used to evaluate fingerprint Individuality, defined as the probability of ran-dom correspondence (PRC) within a tolerance. The evaluation uses Bayesian networks and three fingerprint representations: ridge flow, minutiae, and ridge points. Mixture models of features are used, with parameters estimated using the EM algorithm. Fingerprint PRCs are determined for given numbers of available and matching minutiae. A Bayesian method based on graphical models is introduced to compute fingerprint rarity. The model is able to handle the large number of variables and the complexity of the distributions. Since latent prints are often incomplete and the core point may be missing in the field of view, a machine learning approach based on Gaussian process regression is proposed to predicate the core point. The coordinate system is transformed into standard position based on finding the core point. A graphical model, which takes into account inter-minutia dependencies and minutia i ii confidences, is used to determine evidence probability from which the specific probability of match among n is evaluated. To improve the accuracy of rarity estimation for low qual-ity latent prints, a fully Bayesian treatment allows considering core point uncertainty. The generative model is validated using a goodness-of-fit test. Rarity evaluation is illustrated us-ing several examples, including simple configurations of minutiae, randomly selected latent fingerprints in a database, and a well-known case of erroneous identification