AI threat taxonomy mapping
Mapping that library to NIST’s taxonomy and OWASP AI Exchange is relatively straightforward because of the consensus. However, it does highlight some different approaches to threat identification that are worth exploring.
NIST AI taxonomy mapping
Most entries have a one-to-one mapping with NIST AI 100-2e2023, including data poisoning, backdoor attacks, evasion attacks, model reprogramming, model extraction, prompt injections, data disclosure, and more.
NIST’s taxonomy is more focused on pure Adversarial AI attacks. We include classic cybersecurity threats to aid practical threat modeling and provide more granularity on supply chain attacks.
The NIST taxonomy is also more detailed, with subcategories that have been covered in this book but have been omitted from our threat library. This is intentional; detailed understanding is essential, but too much detail can overwhelm threat modeling without adding value.
Finally, we include...
