Information-theoretic performance evaluation of likelihood-ratio based biometric score fusion under modality selection attacks

Likelihood-ratio based biometric score fusion is gaining much attention, since it maximizes accuracy if a log-likelihood ratio (LLR) is correctly estimated. It can also handle some missing query samples due to adverse physical conditions (e.g. injuries, illness) by setting the corresponding LLRs to 0. In this paper, we refer to the mode that allows missing query samples in such a way as a “modality selection mode”, and clarify a problem with the accuracy in this mode. We firstly propose a “modality selection attack”, which inputs only query samples whose LLRs are more than 0 (i.e. takes an optimal strategy) to impersonate others. We secondly consider the case when both genuine users and impostors take this optimal strategy, and prove information-theoretically that the overall accuracy in this case is “worse” than that in the case when they input all query samples. Specifically, we prove, both theoretically and experimentally, that the KL (Kullback-Leibler) divergence between a genuine distribution of integrated scores and an impostor's one, which can be compared with password entropy, is smaller in the former case. We also show quantitatively to what extent the KL divergence losses.