Model Extraction and Theft COASP Practice Questions (Page 8)

Questions 36–40

1. 36expert · hard

   A company wants to defend its model API against extraction. The security team proposes watermarking the model. The product team asks whether watermarking will prevent extraction. Which response is most accurate?

   • AWatermarking prevents extraction by encrypting the model's parameters
   • BWatermarking prevents extraction by making the model's outputs unusable
   • CWatermarking does not prevent extraction but enables detection and proof of theft
   • DWatermarking prevents extraction by limiting the number of queries
   Reveal Solution Discussion Select an answer first
2. 37application · medium

   A company has deployed a model as a public API. They want to detect if someone is attempting to extract the model. Which monitoring signal is most indicative of an extraction attempt?

   • AA decrease in the model's accuracy over time
   • BA sudden increase in the number of unique users
   • CA high volume of queries from a single client with inputs that are near the model's decision boundary
   • DA steady stream of queries from many different IP addresses
   Reveal Solution Discussion Select an answer first
4. 38application · medium

   A company wants to protect its model from extraction but also needs to provide a high-quality API for legitimate users. They are considering output perturbation. What is the main trade-off they must manage?

   • AReducing the fidelity of extracted models while minimizing the impact on legitimate users' results
   • BEncrypting the model weights while keeping the API accessible
   • CBlocking all queries from suspicious IPs while allowing normal traffic
   • DIncreasing the model's accuracy while reducing its latency
   Reveal Solution Discussion Select an answer first
5. 39application · medium

   A company provides a machine-learning API and wants to defend against model extraction while maintaining a good user experience for legitimate customers. Which combination of defenses best balances security and usability?

   • AReturn only the predicted class and no confidence scores
   • BUse a model with a very large number of parameters to slow down extraction
   • CSet a per-user query limit and round confidence scores
   • DRequire users to submit a business justification for every query
   Reveal Solution Discussion Select an answer first
6. 40expert · hard

   A company's ML team has deployed a fraud-detection model as a public API. They later discover that a competitor has released a model with nearly identical decision boundaries. The company suspects model extraction. They have logs showing the competitor's API key made thousands of queries, but the queries were spread over several months and used varied inputs. The company wants to confirm the theft and take legal action. Which combination of evidence and action is most appropriate?

   • AUse watermarking triggers to test the competitor's model and consult legal counsel if the triggers produce the expected outputs
   • BCompare the model's accuracy on a public benchmark to prove theft
   • CCheck the competitor's API response times to infer the model architecture
   • DAssume the model was stolen and immediately send a cease-and-desist letter
   Reveal Solution Discussion Select an answer first