[2604.04461] DP-OPD: Differentially Private On-Policy Distillation for Language Models

Computer Science > Machine Learning arXiv:2604.04461 (cs) [Submitted on 6 Apr 2026] Title:DP-OPD: Differentially Private On-Policy Distillation for Language Models Authors:Fatemeh Khadem, Sajad Mousavi, Yi Fang, Yuhong Liu View a PDF of the paper titled DP-OPD: Differentially Private On-Policy Distillation for Language Models, by Fatemeh Khadem and 2 other authors View PDF HTML (experimental) Abstract:Large language models (LLMs) are increasingly adapted to proprietary and domain-specific corpora that contain sensitive information, creating a tension between formal privacy guarantees and efficient deployment through model compression. Differential privacy (DP), typically enforced via DP-SGD, provides record-level protection but often incurs substantial utility loss in autoregressive generation, where optimization noise can amplify exposure bias and compounding errors along long rollouts. Existing approaches to private distillation either apply DP-SGD to both teacher and student, worsening computation and the privacy--utility tradeoff, or rely on DP synthetic text generation from a DP-trained teacher, avoiding DP on the student at the cost of DP-optimizing a large teacher and introducing an offline generation pipeline. We propose \textbf{Differentially Private On-Policy Distillation (DP-OPD)}, a synthesis-free framework that enforces privacy solely through DP-SGD on the student while leveraging a frozen teacher to provide dense token-level targets on \emph{student-generated...