In biometric fusion systems, it is common to find the number of available imposter scores being much larger than the number of genuine-user scores. In terms of training a stable fusion classifier, the area under the receiver operating characteristic curve (AUC) could be useful since it is less sensitive to class distributions , , . A direct optimization of this AUC criterion thus becomes a natural choice for fusion classifier design. However, a direct formulation of search based on the AUC criterion would have the incoming data size growing almost exponentially. In this paper, we propose an online learning algorithm to circumvent this computational problem in multi-biometric scores fusion. Since the proposed method involves pairing of data points of opposite classes, an online learning formulation becomes non-trivial. Our empirical results on two publicly available score-level fusion databases show promising potential in terms of verification AUC, Half Total Error Rate, Accuracy, and computational efficiency.