Model Extraction and Theft COASP Practice Questions (Page 7)

Questions 31–35

1. 31expert · hard

   A company has implemented output perturbation and query limiting to defend against model extraction. They now want to add watermarking to their model. What is the primary purpose of adding watermarking?

   • ATo reduce the accuracy of any extracted model
   • BTo provide forensic evidence that can prove model theft in legal proceedings
   • CTo detect extraction attempts in real time
   • DTo slow down the extraction process
   Reveal Solution Discussion Select an answer first
2. 32application · medium

   A company provides a machine-learning API and wants to deter model extraction while maintaining a good user experience for legitimate customers. They decide to add small random noise to the confidence scores returned by the API. What is the primary effect of this defense?

   • AIt degrades the accuracy of a substitute model trained on the noisy outputs
   • BIt detects when an attacker is trying to extract the model
   • CIt increases the model's accuracy on legitimate inputs
   • DIt prevents attackers from sending any queries to the API
   Reveal Solution Discussion Select an answer first
4. 33foundation · easy

   Which method can be used to detect that a model has been extracted and replicated without authorization?

   • AMonitoring the model's API for unusually high query volumes from a single source
   • BReviewing the model's training logs for unauthorized changes
   • CInspecting the model's source code for hidden backdoors
   • DChecking the model's accuracy on the original test set
   Reveal Solution Discussion Select an answer first
5. 34expert · hard

   A company provides a machine-learning API and wants to defend against model extraction. They have a limited engineering budget and must choose between implementing output perturbation or query limiting. Which consideration is most important in making this decision?

   • AOutput perturbation has no impact on legitimate users
   • BOutput perturbation degrades the fidelity of extracted models, while query limiting slows the attack but does not reduce fidelity
   • CQuery limiting is easier to implement than output perturbation
   • DQuery limiting completely prevents model extraction
   Reveal Solution Discussion Select an answer first
6. 35application · medium

   A company exposes a machine-learning model as a public API. To reduce the risk of model extraction, the security team wants to limit the amount of information each response reveals while still providing useful predictions. Which defense is most appropriate?

   • AUse a larger model to make extraction computationally harder
   • BRequire users to agree to a terms-of-service that prohibits model replication
   • CReturn only the top-1 class label and no confidence scores
   • DRound confidence scores to a few decimal places and add small random noise
   Reveal Solution Discussion Select an answer first