The work establishes the first DP regret bound of order O(K^{3/5}) for model-free online RL under general function approximation and the first coverability-based regret bound for batched non-private RL.
Exposing privacy gaps: Membership inference attack on preference data for llm alignment
4 Pith papers cite this work. Polarity classification is still indexing.
4
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
verdicts
UNVERDICTED 4roles
background 1polarities
background 1representative citing papers
PopQuiz Attack infers LLM training data membership by turning examples into quiz questions and measuring answer accuracy, reaching 0.873 average ROC-AUC across six models and outperforming prior methods by 20.6%.
DIBA detects membership of prompts in RLVR training by measuring reward success changes and policy behavioral drift between pre- and post-RLVR model checkpoints.
TabGRAA applies group-relative advantage alignment in an iterative reward-guided post-training loop to improve tabular language model generators on fidelity, utility, and privacy trade-offs across five benchmarks.
citing papers explorer
-
Towards Differentially Private Reinforcement Learning with General Function Approximation
The work establishes the first DP regret bound of order O(K^{3/5}) for model-free online RL under general function approximation and the first coverability-based regret bound for batched non-private RL.
-
Pop Quiz Attack: Black-box Membership Inference Attacks Against Large Language Models
PopQuiz Attack infers LLM training data membership by turning examples into quiz questions and measuring answer accuracy, reaching 0.873 average ROC-AUC across six models and outperforming prior methods by 20.6%.
-
Auditing Data Membership in Reinforcement Learning With Verifiable Rewards
DIBA detects membership of prompts in RLVR training by measuring reward success changes and policy behavioral drift between pre- and post-RLVR model checkpoints.
-
Self-Improving Tabular Language Models via Iterative Reward-Guided Post-Training
TabGRAA applies group-relative advantage alignment in an iterative reward-guided post-training loop to improve tabular language model generators on fidelity, utility, and privacy trade-offs across five benchmarks.