One-class classification in multimodal biometrie authentication

Class imbalance is a major challenge in biometrie authentication, particularly in the context of two-class classification, i.e., distinguishing between genuine users and impostors. Traditional classifiers assume near-balanced class distributions and as such they do not work well when the samples of one class outnumber those of the other. Indeed, class imbalance is a common problem in multimodal biometrics, where typically impostor samples are in the order of 500:1 compared to those of genuine users. In this work, we present the use of one-class classification to enhance multimodal biometric performance in the presence of class imbalance. We consider well-known one-class classifiers, such as the Gaussian Mixture Model, k-Nearest Neighbour, etc in learning the user-specific and user-independent descriptions for the biometric decision inference. We conclude that the user-specific approach is powerful in overcoming the within-class sub-concepts problem, which commonly occurs in multimodal biometrics due to user variation.