User-specific fusion using one-class classification for multimodal biometric systems: Boundary methods

It has been previously shown that the matching performance of a multimodal biometric system can be improved by using user-specific fusion. The objective of this approach is to address the fact that some users are difficult to recognize using some biometric traits, while these traits are highly discriminant for others. Conventional two-class classification methods, when used to design user-specific fusion, often suffer from the problem of limited availability of training data, especially, those of genuine users. In this paper, we propose a user-specific fusion approach, making use of one-class classifiers, known as boundary methods, to avoid the aforementioned problem of the two-class classification approach. We also show that such an approach outperforms others, including the Sum of Scores, the standard SVM, and the one-class SVM, in experiments carried out on the BioSecure DS2 database.