Membership inference
Membership inference: An attack that estimates whether a particular example was present in a model’s training data. Differences in confidence or output can reveal information about people in the training set. Dedicated evaluation and differential privacy help measure and limit the risk.
In italiano: Membership InferenceLegge V · Overfitting, bias e collasso dei modelliDifferences in confidence or output can reveal information about people in the training set. Dedicated evaluation and differential privacy help measure and limit the risk.
Examples
- Comparing members and non-members
- Confidence-score analysis
- Model privacy testing