Model Privacy Attacks COASP Practice Questions (Page 1)

Questions 1–5

1. 1foundation · easy

   What does 'privacy loss' quantify in the context of privacy attacks?

   • AThe amount of additional information an attacker gains about a target from the model's outputs
   • BThe time taken to execute the attack
   • CThe number of queries an attacker sends to the model
   • DThe accuracy of the model on its training data
   Reveal Solution Discussion Select an answer first
2. 2application · medium

   A healthcare startup has trained a model to predict patient readmission risk. A security researcher warns that the model may leak sensitive patient information. The startup wants to assess whether an attacker could determine if a specific patient's record was used in training. Which attack should the security team simulate?

   • AModel inversion attack
   • BModel extraction attack
   • CAttribute inference attack
   • DMembership inference attack
   Reveal Solution Discussion Select an answer first
4. 3foundation · easy

   How does differential privacy help mitigate privacy attacks?

   • ABy adding noise during training so that the model's behavior does not depend too heavily on any single record
   • BBy restricting access to the model's API using authentication
   • CBy encrypting the model's weights at rest
   • DBy removing all sensitive attributes from the training data
   Reveal Solution Discussion Select an answer first
5. 4application · medium

   A government agency publishes a machine learning model trained on citizen data. They want to allow public queries but must reduce the risk of membership inference attacks. Which approach would be most effective while still allowing the model to be queried?

   • ATrain the model with differential privacy
   • BUse a smaller training dataset
   • CLimit the number of queries per user
   • DRemove all categorical features from the training data
   Reveal Solution Discussion Select an answer first
6. 5application · medium

   A company offers a machine learning API that returns detailed confidence scores for all classes. They want to reduce the risk of model extraction. Which defensive measure would be most effective?

   • AUse a deeper neural network
   • BAdd a rate limit of 100 queries per day
   • CIncrease the model's training data
   • DReturn only the top-1 class label instead of full confidence scores
   Reveal Solution Discussion Select an answer first