Date of Award

5-2025

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical Engineering and Computer Science

First Advisor

Michael C. King

Second Advisor

Jignya Patel

Third Advisor

Philip J. Bernhard

Fourth Advisor

Brian A. Lail

Abstract

Security has been a problem for human society for as long as history has been recorded. The identification of people is an ongoing, ancient battle, with a variety of methods that only become more complex with time. The Romans performed censuses, ciphers have been used for thousands of years in the pursuit of security, and in modern day we own identifications and governments keep track of who lives in their country with citizenship and licenses. The question of ”Who are you?” is vital for society to function, which opens up a massive field of potential for how to ask that question and how to lie to it. Biometrics are a modern method of answering that question, relying upon the natural uniqueness of human individuals themselves to identify a person, from fingerprints to irises to the entire face. These methods of identification only work if these biometrics retain their individuality, and continuous research is done to identify biometrics modalities and how to attack them. One such attack is the morph attack, which in some manner blends biometrics together to become a new identity that can pass as any source involved in the creation, completely violating the concept of uniqueness. In our experiment, a publicly available morphing tool for images of faces is used to create presentation attacks of impersonation against two state-of-the-art facial recognition systems to examine how dangerous these attacks might be. By comparing these morphed images against their sources, we found that for one facial recognition model, the attack completely failed, with no morphed faces passing as either source identity, however in the other model, almost all of the morphed faces succeeded in mimicking their sources, performing better than just raw impersonation attempts between similar-looking individuals. This reveals how unpredictable morphing attacks can be, with radically different results with the same exact images across facial recognition, indicating that they are a present threat with low expertise attackers, and difficult to understand what systems may be vulnerable, and how to possibly defend from the idea of morphs as a whole.

Download

Included in

Cybersecurity Commons

Share

COinS