Membership Inference Attacks Quiz

Answers and explanations appear as you take the quiz.

1. What is a membership inference attack trying to determine?
   • Whether a specific record was part of the training set
   • The model's accuracy
   • The model's architecture
   • The size of the test set
2. How can overfitting affect membership-inference exposure in studied models?
   • It proves every training record is identifiable with certainty
   • It removes the need to evaluate attack errors
   • It can increase the behavioral difference between members and comparable nonmembers
   • It makes all outputs independent of training data
3. What technique did the original 2017 Shokri et al. attack rely on?
   • Gradient clipping
   • Shadow models
   • Beam search
   • Tokenization
4. Why can membership inference be harmful even without revealing the record's contents?
   • It slows the model down
   • It changes the model's weights
   • It deletes training data
   • Membership itself can disclose sensitive facts, like a diagnosis
5. What calibration distinguishes the cited LiRA approach from a single naive loss threshold?
   • A guarantee that every record has the same inherent difficulty
   • Only an increased generation temperature
   • Per-example score distributions estimated from in and out models
   • A requirement that the target reveal its entire training dataset
6. What does a properly specified differential-privacy guarantee bound?
   • Changes in output probabilities between neighboring datasets
   • A guarantee that no attack can ever make any correct guess
   • A mandatory deletion of all model outputs
   • Only the raw number of files in a training folder
7. An attack incorrectly flags 10 of 1,000 actual nonmembers as members. What is its false-positive rate on those nonmembers?
   • 10%
   • 0.1%
   • 100%
   • 1%
8. Why does the LiRA paper emphasize true-positive rate at low false-positive rates?
   • False positives are impossible by definition
   • Average accuracy can conceal whether confidently flagged records are reliable membership evidence
   • Every member and nonmember population must have identical sizes
   • The model’s training accuracy directly gives attack precision