Semester

Spring

Date of Graduation

2022

Document Type

Thesis

Degree Type

MS

College

Statler College of Engineering and Mineral Resources

Department

Lane Department of Computer Science and Electrical Engineering

Committee Chair

Nasser Nasrabadi

Committee Co-Chair

Jeremy Dawson

Committee Member

Jeremy Dawson

Committee Member

Matthew Valenti

Abstract

Facial Recognition Systems (FRSs) are a key target for adversaries determined to circumvent security checkpoints. Morph images threaten FRS by presenting as multiple individuals, allowing an adversary to swap identities with another subject. Although morph generation using generative adversarial networks (GANs) results in high-quality morphs without possessing the spatial artifacts caused by landmarkbased methods, there is an apparent loss in identity with standard GAN-based morphing methods. In this thesis, we examine landmark-based and GAN-based morphing methods to fuse the advantages of both methodologies. We propose a novel StyleGAN2 morph generation technique by introducing a landmark enforcement method. Considering this method, we aim to enforce the landmarks of the morph image to represent the spatial average of the landmarks of the bona fide faces.

Loss in visual quality of images projected into the latent space of the StyleGAN2 model reduces the potential quality of the morphs. We compare previous image inversion methods to derive a novel method to improve the latent space representation of an image. To further improve the perceptual quality of the morphs, we examine the noise inputs of our model. Trainability of the noise input is evaluated to learn reconstruction information the latent codes cannot represent. Further exploration of the latent space of our model is conducted using Principal Component Analysis (PCA) to pronounce the effect of the bona fide faces on the morphed latent representation. This work’s contributions include a novel GAN-based morphing method to attack FRS at higher success rates than alternative GAN-based methods. We improve image inversion into the latent space by exploring the model’s noise input while enforcing the balance of latent identities through PCA.

Share

COinS