Exploring Margin Maximization for Biometric Score Fusion Claudio Marrocco, Maria Teresa Ricamato, and Francesco Tortorella DAEIMI - Universit`a degli Studi di Cassino, Cassino, Italy {c.marrocco,mt.ricamato,tortorella}@unicas.it Abstract. Biometric systems are automated methods based on physical or behavioral characteristics of an individual for determining her/his identity. An important aspect of these systems is the reliability against forgery that is surely improved when using multiple sources of biometric information. In such cases combination rules can be applied to fuse the different scores thus obtaining a multibiometric system. In this paper we analyze a method based on margin maximization for building a linear combination of biometric scores. The margin is a cen- tral concept in machine learning research and several theoretical results exist which show that improving the margin on the training set is benefi- cial for the generalization error of an ensemble of classifiers. Experiments performed on real biometric data and comparisons with other commonly employed fusion rules show that a combination based on margin maxi- mization is particularly effective with respect to other established fusion methods. Keywords: multiple classifiers systems, multibiometrics, margins, linear programming. 1 Introduction Biometric systems aim to identify an individual by analyzing some physical characteristics such as fingerprints, iris, face, hand geometry, voice, etc. When dealing with a single source of biometric information, such systems could not ensure a sufficient degree of reliability because of the intrinsic noise affecting biometric data (e.g. a face can be partially occluded by sunglasses, fingerprints could vary due to different temperature and humidity conditions, etc.). Another problem is the vulnerability provided by only one characteristic which can be more easily counterfeited. An effective way to improve the accuracy is to com- bine multiple sources of biometric information: in this way more information is provided to the recognition/verification procedure and the whole process is more reliable and robust against spoofing attempts. Even though such combination could be performed at different levels, the most popular approach is to fuse the scores provided by a pool of classifiers (a.k.a.matchers ) which are obtained by employing different feature extraction and/or matching techniques to different biometric characteristics. It is easy to see that such approach can be arranged in N. da Vitora Lobo et al. (Eds.): SSPR&SPR 2008, LNCS 5342, pp. 674–683, 2008. c  Springer-Verlag Berlin Heidelberg 2008