Model Privacy Attacks COASP Practice Questions (Page 8)

Questions 36–40

1. 36foundation · easy

   What does an attribute inference attack aim to do?

   • ACopy the model's decision function
   • BInfer sensitive attributes of a data record from the model's predictions
   • CDetermine if a record was in the training set
   • DReconstruct entire training images
   Reveal Solution Discussion Select an answer first
2. 37expert · hard

   A tech company trains a recommendation model on user interaction data. They plan to release the model's top-k recommendations for each user as a public dataset for research. A privacy audit shows that the recommendations leak membership information. The company must balance the utility of the dataset with privacy. Which approach is most appropriate?

   • ADelay the publication by six months to make the data less sensitive
   • BPublish only the top-1 recommendation for each user to reduce the amount of information
   • CRemove user IDs from the published dataset
   • DApply differential privacy to the model training and then add noise to the recommendations before publishing
   Reveal Solution Discussion Select an answer first
4. 38application · medium

   A company sells a machine learning model as a paid API. They suspect a competitor is using the API to train a copycat model. The company wants to measure how accurately the competitor's model replicates the original. Which metric would best quantify the fidelity of the extracted model?

   • ATraining accuracy of the extracted model
   • BPrecision of the original model
   • CAgreement rate on a held-out test set
   • DPrivacy loss of the original model
   Reveal Solution Discussion Select an answer first
5. 39expert · hard

   A research lab has trained a model on a small dataset of medical records. They have applied differential privacy with a very small epsilon (e.g., 0.1). They find that the model's accuracy has dropped significantly. They are considering increasing epsilon to 1.0 to improve accuracy. What is the most important trade-off they must consider?

   • AIncreasing epsilon has no effect on privacy or accuracy
   • BIncreasing epsilon improves privacy but reduces accuracy
   • CIncreasing epsilon reduces the privacy guarantee, potentially increasing membership inference risk, but improves model accuracy
   • DIncreasing epsilon reduces both privacy and accuracy
   Reveal Solution Discussion Select an answer first
6. 40application · medium

   A security team is evaluating a membership inference attack against their model. They have a test set of 1000 records, 500 from the training set and 500 from a holdout set. The attack correctly identifies 400 training members and incorrectly labels 100 holdout records as members. What is the attack's precision?

   • A90%
   • B50%
   • C80%
   • D40%
   Reveal Solution Discussion Select an answer first