(A) study on reliable assessment of real-world adversarial robustness of AI models

Abstract

and (3) the Deformable Patch Projection technique, which facilitates physical-world adversarial attacks by modeling each adversarial patch as a deformable mesh and projecting it onto 3D objects using physics simulation. Comprehensive experimental results show that the proposed techniques provide a more reliable assessment of the adversarial robustness of AI models under realistic attack scenarios, which can contribute to the development of more secure and robust AI systems in the future.

(2) the Clean Feature Mixup technique, which further improves the transferability by introducing competition into the optimization process of adversarial examples

Although remarkable AI models have been developed for various computer vision tasks, they still remain vulnerable to adversarial examples, which are subtly modified inputs that lead to incorrect predictions. This vulnerability raises concerns about the security, reliability, and robustness of AI models, especially in real-world applications, and it is therefore important to accurately assess their adversarial robustness in advance. In this regard, this dissertation proposes three novel techniques to improve the attack performance of adversarial examples against black-box target models within realistic adversarial attack scenarios: (1) the Object-based Diverse Input technique, which improves the transferability of adversarial examples by rendering images on randomly sampled 3D objects