Transient Deception: Recent Advances and Trends in Short-Lived Adversarial Examples

Abstract The emergence of adversarial attacks has highlighted the vulnerabilities of deep learning models. This led adversaries to develop novel techniques for generating adversarial examples that are short-lived and imperceptible to humans. These short-lived adversarial examples exploit the temporal aspects of the model and constitute a significant challenge to the robustness and reliability of deep learning systems, as they exploit transient vulnerabilities in the models' decision-making processes. This paper examines the recent trends in short-lived adversarial attacks on machine learning systems. By focusing on the categorisation and taxonomy of these attacks, the distinctions between whitebox vs blackbox, targeted vs untargeted, physical vs digital, and persistent vs short-lived attacks are discussed to provide a comprehensive analysis of the approaches used by adversaries. The paper also explores the increasing sophistication and diversity of short-lived adversarial attack techniques using flickering-based, motion-based, sensor-based, projection-based and gradient-based perturbations to deceive machine learning models. Furthermore, various methods of generating short-lived adversarial examples using these techniques are discussed, providing insights into the evolving aspects of short-lived adversarial attacks and highlighting the need for robust defence mechanisms and other challenges associated with short-lived adversarial examples.