Enhancing Adversarial Robustness: Representation, Ensemble, and Distribution Approaches

Abstract

This thesis seeks to improve adversarial robustness of machine learning models from three important strands including representation learning, ensemble learning and distributional robustness. It offers novel adversarial training frameworks to improve the robustness, while providing a deeper understanding of adversarial vulnerability within the contexts of three aforementioned approaches. This enhanced understanding of adversarial vulnerability paves the way for the development of increasingly robust machine learning models in the future.