46 Spoof Attacks on Multimodal Biometric Systems Zahid Akhtar 1,+ , Sandeep Kale 2 and Nasir Alfarid 3 1 Dept. of Electrical and Electronic Engineering, University of Cagliari, Italy 2 Dept. of Electronic Science, University of Pune, India 3 Cognizant Technology Solutions, India Abstract. Biometrics, referred as the science of recognizing an individual based on his or her physical or behavioral traits, has been widely employed as a security system in the awake of latest security issues. However, recent researches have shown that many biometric traits are vulnerable to spoof attacks. In addition, a latest results have questioned that, contrary to a common claim, multimodal systems can be cracked by spoofing only one trait. Those results were obtained using simulated spoof attacks, under the unrealistic assumption that the spoofed and genuine samples are identical, turned out to be the same outputs. We further investigate this significant security issue, focusing on behavior of fixed and trained score fusion rules, using real spoof attack samples under different spoof attack scenarios. Preliminary empirical results on real biometric systems made up of face, fingerprint and iris with twelve score fusion rules confirm that multimodal biometric systems are not intrinsically robust against spoof attacks as believed so far. In particular, most widely used fixed rules can be less robust, even if the quality of fake biometric trait is low. The false acceptance rate increases substantially under spoof attacks which means that an attacker might wrongly get authenticated by spoofing a subset of traits. In all considered spoofing scenarios, we also found that trained rules are more accurate, flexible and robust against spoof attacks as compare to fixed one. Keywords: Biometrics, Multimodal biometric system, Score fusion rules, Spoof attacks. 1. Introduction Biometrics using fingerprint, face, voice and iris etc. based recognition systems to identify an individual, has been accepted as a legitimate technology. Each biometric trait should pose attributes like uniqueness, and hard to circumvent [1]. Sadly, recent researches have shown that an attacker can lift and replicate the biometric traits, which later can be used to attack on biometric systems [2-4]. As a result, multimodal biometric systems have been proposed to increase the recognition accuracy as well as security against attacks as compared to the unimodal biometric systems that make them up. Several empirical evidences have shown that they are effective to accuracy. It is claimed that multimodal systems are more robust against spoof attacks, since evading several systems is more difficult than evading just one [1]. This claim implies that to evade a multimodal system it is necessary to evade all fused individual systems simultaneously. However, there is no experimental evidence to support this assumption, with the exception of [5], where some evidence was provided that a multimodal biometric system can be fooled by spoofing one of the individual matchers. However, the experiment in [5] was carried out by simulating spoof attacks under the assumption that spoofed and genuine traits are identical which produces same output matching scores, which is not true in real world [6-7]. Hence, it is of great interest to investigate the robustness of multimodal biometric systems against real spoof attacks under more realistic scenarios where a attacker is not able to fabricate a perfect replica of biometric trait. In this work we further contribute to this goal using real spoof attack samples unlike [5]. We analyze the security of multimodal biometric systems by focusing on spoof attacks at the sensor level, which are so far the ones raising the most of interest in the biometric community [8]. They are carried by submitting + Corresponding author. Tel.: + 393294913022 ; fax: + 390706755782. E-mail address: z.momin@diee.unica.it 2011 International Conference on Information and Network Technology IPCSIT vol.4 (2011) © (2011) IACSIT Press, Singapore