Data Hijackers:A Deep Dive into AI Poisoning

This NIST report establishes a standardized taxonomy and terminology for the burgeoning field of adversarial machine learning, categorizing threats to both Predictive and Generative AI. It details a conceptual hierarchy of risks based on attacker goals, such as breaking system availability, violating integrity, or compromising privacy. The text explains specific attack vectors including evasion, where inputs are manipulated at deployment, and poisoning, where training data or models are corrupted. Furthermore, the document addresses privacy concerns like data reconstruction and membership inference, which risk exposing sensitive training information. To help developers and regulators, the report provides an overview of mitigation strategies like adversarial training and differential privacy while acknowledging their inherent trade-offs in accuracy. Ultimately, this framework serves as a foundational guide for assessing and managing the security of AI systems throughout their entire lifecycle.