A differentially private policy gradient method that frames DP clipping as a trust-region choice and demonstrates competitive returns on deep RL and RLHF tasks, with the caveat that harder tasks use weak privacy budgets.
Title resolution pending
1 Pith paper cite this work. Polarity classification is still indexing.
1
Pith paper citing it
fields
cs.LG 1years
2025 1verdicts
CONDITIONAL 1representative citing papers
citing papers explorer
-
Differentially Private Policy Gradient
A differentially private policy gradient method that frames DP clipping as a trust-region choice and demonstrates competitive returns on deep RL and RLHF tasks, with the caveat that harder tasks use weak privacy budgets.