AI securitysrc: MITRE ATLAS
Use Pre-Trained Model
Attackers sometimes use a publicly available AI model as a practice dummy to figure out how to break a more secure, private model they are actually targeting.
Adversaries leverage surrogate models—often pre-trained on similar datasets—to perform black-box testing and generate adversarial examples that are likely to transfer to the target victim model.
Adversaries utilize off-the-shelf pre-trained models as functional proxies to facilitate transfer-based adversarial attacks, exploiting the phenomenon of adversarial transferability to stage and refine malicious inputs against a victim model without requiring direct access to its internal parameters or architecture.