A Systematic Review of Adversarial Attacks: ML Techniques, Classification and Countermeasures

Machine Learning (ML) and Deep Learning (DL)-based technologies have made significant strides in areas such as computer vision, natural language processing, and autonomous systems. Yet they have been applied in high-stake applications, and have been found to be fragile through adversarial attacks-, maliciously crafted small distortions that fool models into making mistakes. From the pioneering work by Szegedy et al. (2014). Adversarial machine learning has since grown apace, including myriad attack strategies and defense approaches. In this systematic review, we study more than 150 peer-reviewed works published during the period of 2014-2025 and provide a holistic taxonomy of attacks based on the knowledge requirement (white-box, gray-box, black-box), attack specificity (targeted vs. untargeted), perturbation nature (pixel-level, spatial, semantic), and persistence in terms of evasion and poisoning. The paper provides a critical assessment of defenses such as adversarial training, gradient masking, input preprocessing, architectural changes, detection and certified defenses. Summary Results show that defense mechanisms evolved enough but no single mechanism is sufficient to achieve total protection against all sort of attacks. Critical research gaps are also discussed on scalability, domain adaptation and robustness-accuracy trade-offs. Several ML methods to reinforce the defense (Hierarchical Ensemble Defense (HED), Distribution-Aware Adversarial Training (DAAT), Self-Supervised robustness Enhancement (SSRE), Neural Architecture Search for Robustness NASR) and method to identify causal features towards increased robustness, Causal Robustness Analysis CRA is introduced, with preliminary experimental evidence. This survey provides the basis for researchers and practitioners interested in understanding, applying, and developing adversarial robustness for ML systems.