Access Type

Open Access Dissertation

Date of Award

January 2022

Degree Type


Degree Name



Computer Science

First Advisor

Dongxiao Zhu


Although deep neural networks (DNNs) have achieved significant advancement in various challenging tasks of computer vision, they are also known to be vulnerable to so-called adversarial attacks. With only imperceptibly small perturbations added to a clean image, adversarial samples can drastically change models’ prediction, resulting in a significant drop in DNN’s performance. This phenomenon poses a serious threat to security-critical applications of DNNs, such as medical imaging, autonomous driving, and surveillance systems. In this dissertation, we present adversarial machine learning approaches for natural image classification and advanced medical imaging systems.

We start by describing our advanced medical imaging systems to tackle the major challenges of on-device deployment: automation, uncertainty, and resource constraint. It is followed by novel unsupervised and semi-supervised robust training schemes to enhance the adversarial robustness of these medical imaging systems. These methods are designed to tackle the unique challenges of defending against adversarial attacks on medical imaging systems and are sufficiently flexible to generalize to various medical imaging modalities and problems. We continue on developing novel training scheme to enhance adversarial robustness of the general DNN based natural image classification models. Based on a unique insight into the predictive behavior of DNNs that they tend to misclassify adversarial samples into the most probable false classes, we propose a new loss function as a drop-in replacement for the cross-entropy loss to improve DNN's adversarial robustness. Specifically, it enlarges the probability gaps between true class and false classes and prevents them from being melted by small perturbations. Finally, we conclude the dissertation by summarizing original contributions and discussing our future work that leverages DNN interpretability constraint on adversarial training to tackle the central machine learning problem of generalization gap.