Robustness of multi-modal biometric systems under realistic spoof attacks against all traits

Spoof attacks consist in submitting fake biometric traits to biometric systems, and are a major threat that can curtail their security. Multi-modal biometric systems are commonly believed to be intrinsically more robust to spoof attacks, but recent works have shown that they can be evaded by spoofing even a single biometric trait. This result was however obtained under the worst-case scenario that the attacker is able to fabricate an exact replica of the genuine biometric trait, which was simulated by assuming that the matching score distribution of fake traits is identical to the one of genuine users. This demands for a more thorough investigation of the robustness of multi-modal biometric systems against realistic spoof attacks, namely under non-worst case scenarios. In this paper we focus on bi-modal systems made up of a face and a fingerprint matcher, whose scores are fused using the well-known sum, product, weighted sum and likelihood ratio (LLR) rules. We evaluate their robustness against realistic spoof attacks obtained by fabricating fake biometric traits. The main goal of our study is to investigate whether a realistic spoof attack against both modalities can allow the attacker to crack the multimodal system. Our results show that even in a realistic, non-worst case scenario, the false acceptance rate (FAR) can remarkably increase.