Improving fusion with one-class classification and boosting in multimodal biometric authentication

Abstract

Class imbalance poses serious difficulties to most standard two-class classifiers, when applied in performing classification in the context of multimodal biometric authentication. In this paper, we propose a system, which exploits the natural capabilities of one-class classifiers in conjunction with the so-called Real AdaBoost to handle the class imbalance problem in biometric systems. Particularly, we propose a decision rule for the fusion of one-class classifiers to effectively use the training data from both classes. By treating this decision rule as the base classifier, the Real AdaBoost is then employed to further improve its performance. An important feature of the proposed system is that it trains the base classifiers with different parameter settings. Hence, it is able to reduce the number of parameters, which are normally set by the user. An empirical evaluation, carried out on the BioSecure DS2 database, demonstrates that the proposed system can achieve a relative performance improvement of 5%, 13%, and 14% as compared to other state-of-the-art techniques, namely the sum of scores, likelihood ratio based score fusion, and support vector machines. © 2014 Springer International Publishing Switzerland.