ConceptPrivate Sparse RLHF

Differentially Private Sparse Reward Estimation

A framework for protecting the privacy of preference data in RLHF that assumes the reward model is governed by only a few relevant features, using sparsity to reduce the dimensionality of differential privacy noise injection and thereby improving the privacy-utility trade-off.

Timeline (last 90 days)

Oct 22

A new study proposes a differentially private sparse reward estimation framework, systematically analyzing the privacy-utility tradeoff in RLHF under the sparsity assumption for the first time.

Unverified50%

All Facts (1)

Unverified

A new study proposes a differentially private sparse reward estimation framework, systematically analyzing the privacy-utility tradeoff in RLHF under the sparsity assumption for the first time.

50%

Source Articles

Differentially Private Sparse Reward Estimation: A New Breakthrough in RLHF Privacy ProtectionOct 22