Biometric fusion: Does modeling correlation really matter?

Sources of information in a multibiometric system are often assumed to be statistically independent in order to simplify the design of the fusion algorithm. However, the independence assumption may not be always valid. In this paper, we analyze whether modeling the dependence between match scores in a multibiometric system has any effect on the fusion performance. Our analysis is based on the likelihood ratio (LR) based fusion framework, which guarantees optimal performance if the match score densities are known. We show that the assumption of independence between matchers has a significant negative impact on the performance of the LR fusion scheme only when (i) the dependence characteristics among genuine match scores is different from that of the impostor scores and (ii) the individual matchers are not very accurate.