MULTIFACTOR USER AUTHENTICATION VIA MULTIMODAL BIOMETRIC DATA

Abstract

A foundational component of computer system security is the design and enforcement of access control measures that serve to both authorize and regulate the use of system resources. A cornerstone of access controls is the implementation of multi-factor user authentication techniques that serve to reliably correlate an on-line identity to a specific user. Biometric methods are increasingly used to verify the "something-youare" component of multi-factor authentication. We propose that using biometrics in varying combinations can serve to improve user authentication by decreasing both false positives and false negatives in the process of user authentication. In this thesis we first examine several biometric methods individually, so as to establish baseline performance, and then in various combinations using fusion scoring with varying weights to determine which combinations yield optimal results. This thesis presents the development and evaluation of a multimodal biometric authentication system integrating face recognition (RGB and thermal images) and voice recognition modules. We utilize the SpeakingFaces dataset, which contains aligned RGB, thermal face images, and audio data of 142 subjects. After developing modules for face and voice recognition, the research thoroughly investigates the individual biometric modalities, evaluating their performance on two dataset sizes derived from the SpeakingFaces dataset: a subset of 20 subjects and a complete dataset of 142 subjects. The results indicate that face recognition modules (RGB and thermal) outperform the voice recognition module, and the performance improves when utilizing larger datasets, emphasizing the significance of extensive training data. Subsequently, the thesis explores the combination of two modules in three different pairings: Face RGB & Voice, Face RGB & Thermal, and Thermal & Voice. The optimal balance of weights between the modules is investigated, revealing that the combination of Face RGB and Thermal images modules achieves the highest performance when more weight is given to the thermal images module. Ultimately, this thesis demonstrates the potential of multimodal biometric authentication systems to provide more accurate and robust user authentication rather than two-modal and single-modal systems. It highlights the importance of combining diverse biometric modalities, optimizing their integration, and the potential utilization of thermal images as a significant factor for performance enhancement. Future research directions include expanding the number of biometric modalities, incorporating larger datasets, and exploring the integration of advanced machine learning techniques, such as deep learning algorithms, to further improve multimodal biometric authentication systems.