Insight into ML security

: adversarial attacks and defences in black-box settings

Abstract

Automatic object detection and recognition (AODR) is one of the most important computer vision techniques utilised in many different domains. Recent advances in deep learning have significantly improved the performance of AODR systems deployed in critical autonomous applications. Nevertheless, adversarial attacks which artificially perturb data to abuse deep learning models are identified as an imminent threat to security and safety critical autonomous applications. This thesis begins with presenting conventional approaches to AODR systems, and subsequently, delves into the emerging threat to AODR systems focusing on adversarial attacks. Finally, the thesis provides a fresh understanding of methods to defend AODR systems from adversarial attacks looking at such attacks from a different perspective.

Date of Award	Dec 2024
Original language	English
Awarding Institution	Queen's University Belfast
Supervisor	Ihsen Alouani (Supervisor) & Niall McLaughlin (Supervisor)