Adversarial Machine Learning Framed by Threat Models, Attack Vectors and Defense Strategies

Machine learning (ML) systems are increasingly integrated into critical real-world applications, from healthcare diagnostics to financial decision-making and network security. However, the growing reliance on ML has exposed inherent vulnerabilities to adversarial attacks—carefully crafted inputs that intentionally deceive models, leading to incorrect predictions and potentially harmful outcomes. These attacks threaten the reliability, integrity, and safety of ML systems, particularly in safety-critical and security-sensitive domains. Understanding adversarial threats and designing robust defenses requires a systematic classification of attack strategies, attacker capabilities, and defensive measures. This study presents a comprehensive taxonomy of adversarial ML threats, categorized by attacker knowledge (white-box, black-box, no-box), attack timing (training-time or inference-time), and adversarial intent (sabotage, espionage, competitive gain, or research purposes). It reviews key adversarial techniques, including evasion, poisoning, and model extraction attacks, and examines state-of-the-art defense mechanisms, organized into input-level, model-level, and detection-based strategies. By providing a structured framework that links threat models to applicable defenses, this work offers researchers and practitioners a practical guide to assess vulnerabilities, prioritize mitigation strategies, and enhance model robustness against evolving adversarial methods. This study contributes to the development of resilient ML systems that maintain both performance and security in complex, adversarial environments.