Model Extraction and Theft COASP Practice Questions (Page 2)

Questions 6–10

1. 6application · medium

   A startup offers a public API that classifies images of retail products. A competitor wants to replicate the model's functionality without access to the training data. The competitor sends thousands of carefully chosen images to the API and records the returned class labels and confidence scores. Which type of attack is the competitor performing?

   • AModel extraction via query-based replication
   • BMembership inference to identify training samples
   • CData poisoning by injecting malicious samples
   • DModel inversion to reconstruct training images
   Reveal Solution Discussion Select an answer first
2. 7expert · hard

   A company has discovered that a competitor's model appears to be a copy of their own. They have logs showing the competitor made many queries to their API, but the queries were spread over a long period and used a variety of inputs. The company wants to confirm the theft. Which action is most likely to provide conclusive evidence?

   • ATest the competitor's model with a set of watermarking triggers that produce unique outputs
   • BAnalyze the response times of the competitor's API
   • CCheck if the competitor's model has the same number of parameters
   • DCompare the models' predictions on a public dataset
   Reveal Solution Discussion Select an answer first
4. 8application · medium

   A company suspects that a competitor has stolen their proprietary image-classification model. They have access to the competitor's public API and want to confirm that the competitor's model is a copy of theirs. Which method is most likely to provide strong evidence of model theft?

   • ACompare the model architectures by observing response times
   • BCheck if the competitor's model has the same training data by sending known samples
   • CMeasure the number of parameters by sending large inputs
   • DSend a set of trigger inputs that produce unique outputs only from the original model
   Reveal Solution Discussion Select an answer first
5. 9foundation · easy

   Which defense against model extraction works by adding noise to the model's output probabilities?

   • AQuery limiting
   • BWatermarking
   • COutput perturbation
   • DModel pruning
   Reveal Solution Discussion Select an answer first
6. 10application · medium

   A security team is reviewing the deployment of a model on an embedded device. They are concerned about an attacker who has physical access to the device and can monitor its power consumption during inference. Which type of extraction attack is this?

   • AMembership inference using the device's logs
   • BAPI-based extraction through the device's network interface
   • CSide-channel extraction via power analysis
   • DModel inversion using the device's output
   Reveal Solution Discussion Select an answer first