Feeding four standard membership-inference scores into an XGBoost classifier yields higher AUC-ROC than the individual attacks on seven datasets for LLMs from 160M to 12B parameters.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.RO 1years
2024 1verdicts
REJECT 1representative citing papers
citing papers explorer
-
EM-MIAs: Enhancing Membership Inference Attacks in Large Language Models through Ensemble Modeling
Feeding four standard membership-inference scores into an XGBoost classifier yields higher AUC-ROC than the individual attacks on seven datasets for LLMs from 160M to 12B parameters.