Model Privacy Attacks COASP Practice Questions (Page 6)

Questions 26–30

1. 26application · medium

   A data science team has trained a model on sensitive customer data. They must choose a defense to reduce membership inference risk while preserving model accuracy. They have decided to use differential privacy. What is the key trade-off they must manage?

   • APrivacy budget has no effect on model accuracy
   • BHigher privacy budget reduces both privacy and accuracy
   • CLower privacy budget reduces privacy protection but increases accuracy
   • DHigher privacy budget (epsilon) reduces privacy protection but increases model accuracy
   Reveal Solution Discussion Select an answer first
2. 27application · medium

   A hospital has trained a model to predict patient mortality. They are required to protect patient privacy. They have decided to use output perturbation by adding noise to the model's predicted probabilities before returning them to users. What is the primary benefit of this approach?

   • AIt increases the model's predictive accuracy
   • BIt guarantees the model will not overfit
   • CIt eliminates the need for differential privacy during training
   • DIt reduces the accuracy of membership inference attacks that rely on confidence scores
   Reveal Solution Discussion Select an answer first
4. 28application · medium

   A research lab trains a generative model on a private dataset of medical images. They expose the model's confidence scores through an API. A security audit wants to test the risk of model inversion. Which evaluation metric is most appropriate to quantify the severity of a successful inversion attack?

   • AThe model's classification accuracy on a test set
   • BThe attack's precision in predicting which images were in the training set
   • CThe number of queries required to perform the attack
   • DThe similarity between reconstructed images and the original training images, measured by a perceptual metric like SSIM
   Reveal Solution Discussion Select an answer first
5. 29foundation · easy

   What is the primary objective of a membership inference attack?

   • ATo reconstruct the exact training data samples
   • BTo determine whether a specific data record was part of the model's training set
   • CTo infer sensitive attributes of a known individual
   • DTo extract the model's decision boundary
   Reveal Solution Discussion Select an answer first
6. 30application · medium

   A facial recognition model is deployed as an API that returns confidence scores for each identity class. A security team is concerned that an attacker could reconstruct a recognizable face image from the model's output. Which defensive technique would most directly reduce the risk of this reconstruction?

   • ARestrict the API to return only the top-1 class label
   • BUse a larger training dataset
   • CApply L2 regularization during training
   • DAdd noise to the output confidence scores
   Reveal Solution Discussion Select an answer first