Which type of attack targets a deployed AI model by querying it with inputs designed to extract information about its internal architecture, hyperparameters, or training data statistics?

Question

Accepted Answer

A. Model extraction (model stealing) attack. Model extraction attacks send many queries to a model API to reconstruct an approximate copy of the model or extract information about its architecture and training. The SecAI+ objectives list protecting AI models as a core security control requirement. Evasion attacks manipulate specific inputs to cause misclassification without trying to reconstruct the model. Backdoor attacks embed hidden triggers during training.

Which type of attack targets a deployed AI model by querying it with inputs designed to extract information about its internal architecture, hyperparameters, or training data statistics?

Related Questions

Discussion