This dissertation analyzed adversarial machine learning and its applications in computer vision and biometrics.
The first part of the dissertation examined the adversarial susceptibility of deep learning models, providing an empirical analysis on the extent of vulnerability by proposing two adversarial attacks that explored the geometric and frequency-domain characteristics of inputs to manipulate deep decisions. Inspired by theoretical findings, a reliable and practical defense against adversarial examples was formalized to make ensembles robust. The second part of the dissertation harnessed adversarial learning to improve the generalization and performance of deep networks in discriminative and generative tasks. Several models for biometric identification were developed, including fingerprint distortion rectification and latent fingerprint reconstruction. A ridge reconstruction model was developed based on generative adversarial networks that estimate the missing ridge information in latent fingerprints. This report also discusses various aspects of the methods and applications developed that can be improved in the future or incorporated in other applications to improve their performance. 57 figures, 36 tables, and 289 references
810 Seventh Street NW, Washington, DC 20531, United States