Abstract:
User authentication is a fundamental requirement of any role-based access control
system, governing both physical and digital access to organizational resources, and
the related security and privacy of data and transactional meta-data. In our paper
we review methods of authentication based on biometric characteristics such as
fingerprint, retina, hand geometry, face geometry, face thermogram, voice and handwriting.
We replicated recent work on multimodal biometric authentication, using
aligned streams of audio and video data, and examined obfuscation techniques that
could be used to undermine confidence in those techniques.
Based on this experience, we designed and implemented a system for combined face
and voice authentication using the open-access SpeakingFaces dataset. Vocal features
are extracted using Mel-Frequency Cepstral Coefficients (MFCCs), and facial features
are obtained with Local Binary Patterns (LBPs). Face and voice identification are
performed using image similarity with the Euclidean distances metric and Gaussian
Mixture Model (GMM) respectively, and in turn combined into a single multimodal
system using matching scores fusion.
The multimodal biometric authentication system was assessed using open-source
data from Georgia Tech Face Database and the DARPA TIMIT Acoustic-Phonetic
Continuous Speech Corpus. The confidentiality of the face and voice recognition system
was analyzed with several scenarios using spoofing of facial features, imitation of
voice features, combined spoofing, and no spoofing scenarios.This project successfully
replicated the published work, improved the computational performance and demonstrated
that the ranking based model of multimodal biometric system is more resilient
than a threshold-based system. Reported weaknesses of the prior works were used to
improve the performance of our multimodal biometric authentication system.