Adversarial Machine Learning in Cybersecurity: A Review on Defending Against AI-Driven Attacks
Mitra Penmetsa, Jayakeshav Reddy Bhumireddy, Rajiv Chalasani, Srikanth Reddy Vangala, Ram Mohan Polam, Bhavana Kamarthapu · European Journal of Applied Science Engineering and Technology · 2025
The application of artificial intelligence (AI) and machine learning (ML) in cybersecurity has revolutionized the capacity to identify, prevent, and react to more complex cyber threats. However, as ML models become central to defense mechanisms, adversarial attacks designed to deceive these models have emerging as a major problem. Adversarial Machine Learning (AML) focuses on how attackers manipulate data inputs to exploit vulnerabilities in ML systems, leading to misclassification, data breaches, and system failures. This article gives a detailed study of adversarial attacks against ML models in cybersecurity, categorizing them into evasion, poisoning, and inference attacks based on their objectives and methodologies. Furthermore, it examines threat models, adversary knowledge levels (white-box, black-box, gray-box), and common attack techniques such as FGSM, Deep Fool, and model extraction. Adversarial training is one example of defensive techniques, defensive distillation, and input preprocessing are also discussed to highlight how researchers and practitioners are working to Improve the robustness of AI-powered security systems. This study aims to offer a detailed description of the adversarial environment in cybersecurity, providing as a reference for future studies and actual implementations of strong and secure ML systems.