Date of Graduation
Statler College of Engineering and Mineral Resources
Lane Department of Computer Science and Electrical Engineering
Facial recognition systems play a vital role in our everyday lives. We rely on this technology from menial tasks to issues as vital as national security. While strides have been made over the past ten years to improve facial recognition systems, morphed face images are a viable threat to the reliability of these systems. Morphed images are generated by combining the face images of two subjects. The resulting morphed face shares the likeness of the contributing subjects, confusing both humans and face verification algorithms. This vulnerability has grave consequences for facial recognition systems used on international borders or for law enforcement purposes. To detect these morph images, high-quality data must be generated to improve deep morph detectors.
In this work, high-quality morph images are generated to fool these deep morph detection algorithms. This work creates some of the most challenging large-scale morphed datasets to date. This is done in three ways. First, rather than utilizing typical datasets used for face morphing found in literature, we generate morphed data from underrepresented groups of individuals to further increase the difficulty of morphs. Second, we generate morph subjects using a wavelet decomposition blending technique to generate morph images that may perform better than typical landmark morphs while creating morph images that may appear different to detectors than what is seen in literature. Third, we apply adversarial perturbation to the morph images to further increase their attack capability on morph detectors. Using these techniques, the generated morph datasets are highly successful at fooling facial recognition systems into erroneously classifying a morph as a bona fide subject.
O'Haire, Kelsey Lynn, "Generation of High Performing Morph Datasets" (2022). Graduate Theses, Dissertations, and Problem Reports. 11287.