Modeling per-token loss correlations with a multivariate Gaussian improves LiRA membership inference on sequence models, often by an order of magnitude in TPR at low FPR.
What does it mean for a language model to preserve privacy?
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.CR 1years
2025 1verdicts
CONDITIONAL 1roles
background 1polarities
background 1representative citing papers
citing papers explorer
-
Membership Inference Attacks on Sequence Models
Modeling per-token loss correlations with a multivariate Gaussian improves LiRA membership inference on sequence models, often by an order of magnitude in TPR at low FPR.