Model Extraction and Theft COASP Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   An attacker is extracting a remote sentiment-analysis model. The API returns only the predicted class (positive/negative) and not confidence scores. The attacker wants to approximate the model's decision boundary. Which query strategy is most effective under this constraint?

   • ASend identical inputs multiple times to average the outputs
   • BSend inputs with high confidence to extract the model's parameters directly
   • CSend random inputs and record the labels to train a model with noisy labels
   • DSend inputs near the suspected boundary and use the binary labels to refine the boundary
   Reveal Solution Discussion Select an answer first
2. 27foundation · easy

   In a model extraction attack, what does the attacker typically have access to?

   • AThe model's weights and biases
   • BThe model's source code and architecture
   • CThe model's inference API, which returns predictions for submitted inputs
   • DThe model's training dataset and hyperparameters
   Reveal Solution Discussion Select an answer first
4. 28application · medium

   A security team is reviewing logs and notices that a client IP sends a high volume of queries to a model API, with inputs that are slightly perturbed versions of each other. The team suspects model extraction. Which detection method is most effective in this scenario?

   • AMonitor query rate and input similarity to flag extraction-like patterns
   • BReview the client's billing history for unusual payment methods
   • CCheck the client's IP address against a threat-intelligence feed
   • DCompare the client's queries to the training data distribution
   Reveal Solution Discussion Select an answer first
5. 29application · medium

   A company has deployed a model that is a key differentiator for its product. The security team wants to detect if an attacker has successfully extracted the model. Which approach is most practical?

   • ACompare the performance of the company's model to public models on a benchmark
   • BEmbed a watermark and periodically query suspicious third-party APIs with the trigger
   • CUse a honeypot model that returns fake predictions to unknown users
   • DMonitor the company's own API logs for high query volumes
   Reveal Solution Discussion Select an answer first
6. 30foundation · easy

   Which extraction attack vector relies on observing physical characteristics of the hardware running the model?

   • AAPI-based extraction
   • BSocial engineering extraction
   • CQuery-based extraction
   • DSide-channel extraction
   Reveal Solution Discussion Select an answer first