Privacy-Preserving Reinforcement Learning from Human Feedback via Decoupled Reward Modeling

• URL: http://arxiv.org/abs/2603.22563v1
• Date: Mon, 23 Mar 2026 20:45:17 GMT
• Title: Privacy-Preserving Reinforcement Learning from Human Feedback via Decoupled Reward Modeling
• Abstract summary: We propose a privacy-preserving framework that imposes differential privacy only on reward learning.<n>We show that privacy contributes an additional additive term beyond the usual non-private statistical error.<n>Experiments on the Anthropic HH-RLHF dataset show stronger private alignment performance than existing differentially private baseline methods.
• Score: 2.805986764620217
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Preference-based fine-tuning has become an important component in training large language models, and the data used at this stage may contain sensitive user information. A central question is how to design a differentially private pipeline that is well suited to the distinct structure of reinforcement learning from human feedback. We propose a privacy-preserving framework that imposes differential privacy only on reward learning and derives the final policy from the resulting private reward model. Theoretically, we study the suboptimality gap and show that privacy contributes an additional additive term beyond the usual non-private statistical error. We also establish a minimax lower bound and show that the dominant term changes with sample size and privacy level, which in turn characterizes regimes in which the upper bound is rate-optimal up to logarithmic factors. Empirically, synthetic experiments confirm the scaling predicted by the theory, and experiments on the Anthropic HH-RLHF dataset using the Gemma-2B-IT model show stronger private alignment performance than existing differentially private baseline methods across privacy budgets.

Related papers

• On the MIA Vulnerability Gap Between Private GANs and Diffusion Models [51.53790101362898]
  Generative Adversarial Networks (GANs) and diffusion models have emerged as leading approaches for high-quality image synthesis.<n>We present the first unified theoretical and empirical analysis of the privacy risks faced by differentially private generative models.
  arXiv  Detail & Related papers  (2025-09-03T14:18:22Z)
• Model Agnostic Differentially Private Causal Inference [16.50501378936487]
  Estimating causal effects from observational data is essential in medicine, economics and social sciences.<n>We propose a general, model-agnostic framework for differentially private estimation of average treatment effects.
  arXiv  Detail & Related papers  (2025-05-26T07:00:37Z)
• Multi-Objective Optimization for Privacy-Utility Balance in Differentially Private Federated Learning [12.278668095136098]
  Federated learning (FL) enables collaborative model training across distributed clients without sharing raw data.<n>We propose an adaptive clipping mechanism that dynamically adjusts the clipping norm using a multi-objective optimization framework.<n>Our results show that adaptive clipping consistently outperforms fixed-clipping baselines, achieving improved accuracy under the same privacy constraints.
  arXiv  Detail & Related papers  (2025-03-27T04:57:05Z)
• Differentially Private Random Feature Model [47.35176457481132]
  We produce a differentially private random feature model for privacy-preserving kernel machines.<n>We show that our method preserves privacy and derive a generalization error bound for the method.
  arXiv  Detail & Related papers  (2024-12-06T05:31:08Z)
• Initialization Matters: Privacy-Utility Analysis of Overparameterized Neural Networks [72.51255282371805]
  We prove a privacy bound for the KL divergence between model distributions on worst-case neighboring datasets. We find that this KL privacy bound is largely determined by the expected squared gradient norm relative to model parameters during training.
  arXiv  Detail & Related papers  (2023-10-31T16:13:22Z)
• An Empirical Analysis of Fairness Notions under Differential Privacy [3.3748750222488657]
  We show how different fairness notions, belonging to distinct classes of statistical fairness criteria, are impacted when one selects a model architecture suitable for DP-SGD. These findings challenge the understanding that differential privacy will necessarily exacerbate unfairness in deep learning models trained on biased datasets.
  arXiv  Detail & Related papers  (2023-02-06T16:29:50Z)
• On the Statistical Complexity of Estimation and Testing under Privacy Constraints [17.04261371990489]
  We show how to characterize the power of a statistical test under differential privacy in a plug-and-play fashion. We show that maintaining privacy results in a noticeable reduction in performance only when the level of privacy protection is very high. Finally, we demonstrate that the DP-SGLD algorithm, a private convex solver, can be employed for maximum likelihood estimation with a high degree of confidence.
  arXiv  Detail & Related papers  (2022-10-05T12:55:53Z)
• Don't Generate Me: Training Differentially Private Generative Models with Sinkhorn Divergence [73.14373832423156]
  We propose DP-Sinkhorn, a novel optimal transport-based generative method for learning data distributions from private data with differential privacy. Unlike existing approaches for training differentially private generative models, we do not rely on adversarial objectives.
  arXiv  Detail & Related papers  (2021-11-01T18:10:21Z)
• Understanding Clipping for Federated Learning: Convergence and Client-Level Differential Privacy [67.4471689755097]
  This paper empirically demonstrates that the clipped FedAvg can perform surprisingly well even with substantial data heterogeneity. We provide the convergence analysis of a differential private (DP) FedAvg algorithm and highlight the relationship between clipping bias and the distribution of the clients' updates.
  arXiv  Detail & Related papers  (2021-06-25T14:47:19Z)
• PEARL: Data Synthesis via Private Embeddings and Adversarial Reconstruction Learning [1.8692254863855962]
  We propose a new framework of data using deep generative models in a differentially private manner. Within our framework, sensitive data are sanitized with rigorous privacy guarantees in a one-shot fashion. Our proposal has theoretical guarantees of performance, and empirical evaluations on multiple datasets show that our approach outperforms other methods at reasonable levels of privacy.
  arXiv  Detail & Related papers  (2021-06-08T18:00:01Z)
• Robustness Threats of Differential Privacy [70.818129585404]
  We experimentally demonstrate that networks, trained with differential privacy, in some settings might be even more vulnerable in comparison to non-private versions. We study how the main ingredients of differentially private neural networks training, such as gradient clipping and noise addition, affect the robustness of the model.
  arXiv  Detail & Related papers  (2020-12-14T18:59:24Z)
• Differentially Private Federated Learning with Laplacian Smoothing [72.85272874099644]
  Federated learning aims to protect data privacy by collaboratively learning a model without sharing private data among users. An adversary may still be able to infer the private training data by attacking the released model. Differential privacy provides a statistical protection against such attacks at the price of significantly degrading the accuracy or utility of the trained models.
  arXiv  Detail & Related papers  (2020-05-01T04:28:38Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.