Journal of Artificial Intelligence, Machine Learning and Soft Computing

Abstract

Machine learning (ML) has revolutionized various sectors including healthcare, finance, autonomous systems, and cybersecurity. However, real-world deployment exposes models to uncertainties, noisy inputs, and malicious adversarial attacks. Robust machine learning focuses on enhancing model stability against noise and perturbations, while adversarial machine learning studies deliberate attacks that exploit model vulnerabilities. This paper provides a comprehensive review of robust and adversarial ML, covering recent advancements, threat models, attack and defense mechanisms, evaluation metrics, and practical applications. Additionally, we explore the challenges of achieving both robustness and accuracy and highlight future research directions. A detailed discussion of defense strategies, including adversarial training, certified robustness, and robust optimization, is presented, along with comparisons of current methodologies.

Keywords: Robust Machine Learning, Adversarial Attacks, Adversarial Training, Certified Robustness, Deep Learning Security, Threat Models, Perturbation Analysis