ADAPTIVE BIOMETRIC AUTHENTICATION USING NONLINEAR MAPPINGS ON QUALITY MEASURES AND VERIFICATION SCORES Jinyu Zuo, Francesco Nicolo, Natalia A. Schmid West Virginia University Department of CSEE, Morgantown, WV Harry Wechsler George Mason University Department of CS, Fairfax, VA ABSTRACT Three methods to improve the performance of biometric matchers based on vectors of quality measures associated with biometric samples are described. The first two meth- ods select samples and matching scores based on predicted values of Quality of Sample (QS) index (defined here as d- prime) and Confidence in matching Scores (CS), respectively. The third method treats quality measures as weak but useful features for discrimination between genuine and imposter matching scores. The unifying theme for the three methods consists of a nonlinear mapping between quality measures and the predicted values of QS, CS, and combined quality measures and matching scores, respectively. The proposed methodology is generic and is suitable for any biometric modality. The experimental results reported show significant performance improvements for all the three methods when applied to iris biometrics. Index Terms— Quality factors, pattern recognition 1. INTRODUCTION The most common definition of biometric sample quality is at the image or signal level. A quality checking block is in- troduced into every biometric system to ensure that enrolled image/signals have sufficient quality to be further processed. Many recent biometric systems extract a vector of quality measures. The components of a vector of quality measures, however, rarely carry equal weight in terms of their relation- ship to the performance of the matcher. In practical appli- cations (such as US Visit program), it is required to keep a single biometric quality measure in order to decide if biomet- ric samples are suitable for further processing and matching. Research questions should thus be concerned with (i) what quality measures to use; (ii) how to combine multiple quality measures into a single quality index without losing the infor- mation that the vector of quality measures contains; and (iii) how to use this vector to improve performance of biometric systems? This paper addresses all those questions. Most of the quality based matchers described in the liter- ature involve biometric sample quality at the matching stage by concatenating matching scores due to the original matcher and quality measures. These matchers are known as Q-stack classifiers [1, 2]. In spite of the fundamental theory presented in these works in support of Q-stack classifiers, the improve- ment of performance is marginal, if at all (see [3], [1]). More noticeable improvements are reported for Q-stack classifiers operating on multiple algorithms or multiple matchers [3, 2]. This paper suggests several methods on the use of bio- metric sample quality to improve the performance of a single matcher. It targets two main applications for quality mea- sures: 1) to improve performance of a matcher by predicting its QS index or CS score and using them to decide if the un- derlying biometric sample should be retained or discarded 2) to design a nonlinear matcher that treats a vector of quality measures as a set of weak features. The remainder of the paper is organized as follows. Sec. 2 describes the three proposed methods. Sec. 3 describes the data sets used and presents experimental results. Sec. 4 summarizes the contributions. 2. METHODOLOGY This paper advocates the use of predictive tools for the design of quality enhanced matchers. The tools proposed predict a set of quality measures and scores. In each of the three meth- ods described below, the functional relationship between vec- tors of quality measures and the predicted (estimated) mea- sures is not known and has to be modeled. The modeling problem is stated as a multivariate regression problem: Y = f (X 1 ,...,X K ), (1) where f is a multivariate adaptive mapping, variable Y is the estimated output variable, i.e. that characterizes the overall quality, confidence in matching score, or quality enhanced decision, and X 1 ,...,X K is a vector of K input (predictive) variables, such as a vector of quality measures for a biomet- ric sample or a concatenated vector of quality measures and matching scores. Since the true relationship between the in- put and output variables is not known, it is estimated using a set of labeled training data. The multivariate adaptive map- ping f ( · ) can be implemented using a variety of multivariate functions and systems. The results reported are obtained us- ing a feed forward neural network (FFNN) (see Sec. 3).