07/01/2023
MITRE ATLAS (Adversarial Landscape for Artificial-Intelligence Systems):
✔️ includes adversary and for machine learning (ML) systems.
✔️ is based on real-world observations from red teams, cyber security working groups and academic research.
✔️ is modeled after the MITRE ATT&CK
Examples of specific attacks:
📌Model Evasion – attacker identifies query to perform to get a desired outcome after observing output of the ML system for certain types of input (inference stage)
📌Functional Extraction – obtaining model functionally through iterative querying (inference stage)
📌Model Poisoning – attacker can “reprogram” the ML system to perform a new undesired task or have a new undesired output (train phase)
📌Model Inversion – attacker recovers the features used to train the model (inference stage)
https://atlas.mitre.org/
&CK