Abstract:
This thesis addresses the problem of biometrics security and accuracy through the
fusion of fingerprint and face images. The Biometrics community in recent years
came up with different approaches to improve the accuracy and security of systems.
Multimodal authentication has attracted a lot of attention because of its advantage
over single biometrics matchers. Even though efforts were made to improve these
systems, they are, however, still vulnerable to security threats such as spoofing
where a forged biometric copy and/or artificially recreated biometric data (which
maybe legitimate) may be used to spoof the system. Multimodal biometric systems
overcome various limitations of uni-modal biometric systems, such as nonuniversality,
lower false acceptance, and higher genuine acceptance rates. We thus
propose a framework, based on Bayesian Network, which is tailored to deal with the
fusion of fingerprint and face. In our proposed framework, face and fingerprint data
are fused at Feature level, using probabilistic inference to map posterior probability
distributions of unobserved variable to make a basic decision of authentication. We
present our proposed framework for fusing bimodal biometrics systems by
combining face and fingerprint images to better the performance of biometrics
security system. In this research we wanted to exploit prior knowledge as much as
possible hence our use of the Bayesian Network to provide reasoning to our
framework. Although Bayesian Networks have been used before, our focus in this
study is to fully exploit the graphical structure of Bayesian Networks and explicitly
model their statistical dependencies between relevant variables per sample
measurement. The evaluation of the proposed framework used multimodal
chimerical databases formed from publicly available databases. The evaluation of
our framework shows an improved performance of 0.03 Error Equal Rate over the
performance of the face or finger biometric modalities when each is implemented.
The Error Equal Rate for Face and Finger biometric modalities were 0.31 and 0.28,
respectively.
