Adversarial Machine Learning: A 20-Year Survey of Attacks, Defenses, and Standards

Adversarial Machine Learning (AML) presents a significant barrier to the large-scale deployment of Artificial Intelligence (AI) in safety-critical environments. While early research focused on algorithmic robustness, the field has evolved into a complex intersection of security, assurance, and policy. This paper presents a comprehensive, multidisciplinary survey of the AML landscape, covering over 250 peer-reviewed contributions. We implement a lifecycle-oriented taxonomy that maps attack vectors and defense mechanisms to specific stages of the AI pipeline from data collection to deployment, expanding the traditional Confidentiality, Integrity, and Availability (CIA) triad to include Governance and Regulation.We identify critical research gaps, including certified robustness for Natural Language Processing (NLP), and emerging threats in Generative AI. To ground these theoretical insights in practice, we analyze five domain-specific case studies: Autonomous Vehicles, Medical AI, Financial Systems, Natural Language Processing (NLP), and the Internet of Things (IoT). Uniquely, this survey bridges the gap between academic literature and industrial practice by mapping technical AML findings to emerging standards, including the NIST AI Risk Management Framework (RMF), MITRE ATLAS, and ISO/IEC 42001. We conclude by providing a roadmap for researchers, practitioners, and regulators to build verifiable, trustworthy, and compliant AI systems.