@inproceedings{ngong-etal-2025-differentially,
title = "Differentially Private Learning Needs Better Model Initialization and Self-Distillation",
author = "Ngong, Ivoline C. and
Near, Joseph and
Mireshghallah, Niloofar",
editor = "Chiruzzo, Luis and
Ritter, Alan and
Wang, Lu",
booktitle = "Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
month = apr,
year = "2025",
address = "Albuquerque, New Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.naacl-long.455/",
doi = "10.18653/v1/2025.naacl-long.455",
pages = "9009--9027",
ISBN = "979-8-89176-189-6",
abstract = "Differentially private SGD (DPSGD) enables privacy-preserving training of language models, but often reduces utility, diversity, and linguistic quality. We introduce DPRefine, a three-phase method that initializes a model using data synthesis from a small pre-trained LM with rigorous filtering, applies DP finetuning on private data, and performs self-distillation to refine outputs. This approach significantly outperforms vanilla DPSGD, with AlpacaEval preferring DPRefine{'}s generations in 78.38{\%} of cases across all datasets and metrics, while also demonstrating substantial improvements in lexical diversity, achieving 85.31{\%} in MSTTR and 86.82{\%} in Jaccard similarity. Our fine-grained analysis reveals that DPRefine reduces linguistic errors in generated text by 84{\%}, mitigating grammar errors, spelling mistakes, and missing punctuation commonly associated with DPSGD. It also reduces inconsistencies present in non-private models, such as fabricated details and misattributed quotes. We find that small models like GPT-2 and T5 are effective for initialization and distillation, highlighting their potential in enabling scalable and efficient deployment of high-performing, privacy-preserving language models with improved linguistic quality and consistency."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ngong-etal-2025-differentially">
<titleInfo>
<title>Differentially Private Learning Needs Better Model Initialization and Self-Distillation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Ivoline</namePart>
<namePart type="given">C</namePart>
<namePart type="family">Ngong</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Joseph</namePart>
<namePart type="family">Near</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Niloofar</namePart>
<namePart type="family">Mireshghallah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-04</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Luis</namePart>
<namePart type="family">Chiruzzo</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Alan</namePart>
<namePart type="family">Ritter</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Lu</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Albuquerque, New Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-189-6</identifier>
</relatedItem>
<abstract>Differentially private SGD (DPSGD) enables privacy-preserving training of language models, but often reduces utility, diversity, and linguistic quality. We introduce DPRefine, a three-phase method that initializes a model using data synthesis from a small pre-trained LM with rigorous filtering, applies DP finetuning on private data, and performs self-distillation to refine outputs. This approach significantly outperforms vanilla DPSGD, with AlpacaEval preferring DPRefine’s generations in 78.38% of cases across all datasets and metrics, while also demonstrating substantial improvements in lexical diversity, achieving 85.31% in MSTTR and 86.82% in Jaccard similarity. Our fine-grained analysis reveals that DPRefine reduces linguistic errors in generated text by 84%, mitigating grammar errors, spelling mistakes, and missing punctuation commonly associated with DPSGD. It also reduces inconsistencies present in non-private models, such as fabricated details and misattributed quotes. We find that small models like GPT-2 and T5 are effective for initialization and distillation, highlighting their potential in enabling scalable and efficient deployment of high-performing, privacy-preserving language models with improved linguistic quality and consistency.</abstract>
<identifier type="citekey">ngong-etal-2025-differentially</identifier>
<identifier type="doi">10.18653/v1/2025.naacl-long.455</identifier>
<location>
<url>https://aclanthology.org/2025.naacl-long.455/</url>
</location>
<part>
<date>2025-04</date>
<extent unit="page">
<start>9009</start>
<end>9027</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Differentially Private Learning Needs Better Model Initialization and Self-Distillation
%A Ngong, Ivoline C.
%A Near, Joseph
%A Mireshghallah, Niloofar
%Y Chiruzzo, Luis
%Y Ritter, Alan
%Y Wang, Lu
%S Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)
%D 2025
%8 April
%I Association for Computational Linguistics
%C Albuquerque, New Mexico
%@ 979-8-89176-189-6
%F ngong-etal-2025-differentially
%X Differentially private SGD (DPSGD) enables privacy-preserving training of language models, but often reduces utility, diversity, and linguistic quality. We introduce DPRefine, a three-phase method that initializes a model using data synthesis from a small pre-trained LM with rigorous filtering, applies DP finetuning on private data, and performs self-distillation to refine outputs. This approach significantly outperforms vanilla DPSGD, with AlpacaEval preferring DPRefine’s generations in 78.38% of cases across all datasets and metrics, while also demonstrating substantial improvements in lexical diversity, achieving 85.31% in MSTTR and 86.82% in Jaccard similarity. Our fine-grained analysis reveals that DPRefine reduces linguistic errors in generated text by 84%, mitigating grammar errors, spelling mistakes, and missing punctuation commonly associated with DPSGD. It also reduces inconsistencies present in non-private models, such as fabricated details and misattributed quotes. We find that small models like GPT-2 and T5 are effective for initialization and distillation, highlighting their potential in enabling scalable and efficient deployment of high-performing, privacy-preserving language models with improved linguistic quality and consistency.
%R 10.18653/v1/2025.naacl-long.455
%U https://aclanthology.org/2025.naacl-long.455/
%U https://doi.org/10.18653/v1/2025.naacl-long.455
%P 9009-9027
Markdown (Informal)
[Differentially Private Learning Needs Better Model Initialization and Self-Distillation](https://aclanthology.org/2025.naacl-long.455/) (Ngong et al., NAACL 2025)
ACL