Membership Inference
A privacy attack and evaluation method that determines whether a given data sample belongs to a model's training set, commonly used for copyright and privacy auditing.
Timeline (last 90 days)
Under the counterfactual reference, the area under the ROC curve (AUC) of membership inference reached 0.767.
When a single threshold is used across domains, only 3.8% of member documents can be captured at a 1% false positive rate.
When pooled across domains, the reference-free attack's AUC gap from random guessing does not exceed 0.07.
Membership inference remains unreliable at the document level, but set-level detection is feasible.
All Facts (4)
Under the counterfactual reference, the area under the ROC curve (AUC) of membership inference reached 0.767.
50%UnverifiedWhen a single threshold is used across domains, only 3.8% of member documents can be captured at a 1% false positive rate.
50%UnverifiedWhen pooled across domains, the reference-free attack's AUC gap from random guessing does not exceed 0.07.
50%UnverifiedMembership inference remains unreliable at the document level, but set-level detection is feasible.
50%