@inproceedings{flemings-etal-2024-differentially,
title = "Differentially Private Next-Token Prediction of Large Language Models",
author = "Flemings, James and
Razaviyayn, Meisam and
Annavaram, Murali",
editor = "Duh, Kevin and
Gomez, Helena and
Bethard, Steven",
booktitle = "Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)",
month = jun,
year = "2024",
address = "Mexico City, Mexico",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2024.naacl-long.247",
pages = "4390--4404",
abstract = "Ensuring the privacy of Large Language Models (LLMs) is becoming increasingly important. The most widely adopted technique to accomplish this is DP-SGD, which trains a model to guarantee Differential Privacy (DP). However, DP-SGD overestimates an adversary{'}s capabilities in having white box access to the model and, as a result, causes longer training times and larger memory usage than SGD. On the other hand, commercial LLM deployments are predominantly cloud-based; hence, adversarial access to LLMs is black-box. Motivated by these observations, we present Private Mixing of Ensemble Distributions (PMixED): a private prediction protocol for next-token prediction that utilizes the inherent stochasticity of next-token sampling and a public model to achieve Differential Privacy. We formalize this by introducing RD-mollifers which project each of the model{'}s output distribution from an ensemble of fine-tuned LLMs onto a set around a public LLM{'}s output distribution, then average the projected distributions and sample from it. Unlike DP-SGD which needs to consider the model architecture during training, PMixED is model agnostic, which makes PMixED a very appealing solution for current deployments. Our results show that PMixED achieves a stronger privacy guarantee than sample-level privacy and outperforms DP-SGD for privacy $\epsilon = 8$ on large-scale datasets. Thus, PMixED offers a practical alternative to DP training methods for achieving strong generative utility without compromising privacy.",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="flemings-etal-2024-differentially">
<titleInfo>
<title>Differentially Private Next-Token Prediction of Large Language Models</title>
</titleInfo>
<name type="personal">
<namePart type="given">James</namePart>
<namePart type="family">Flemings</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Meisam</namePart>
<namePart type="family">Razaviyayn</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Murali</namePart>
<namePart type="family">Annavaram</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2024-06</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)</title>
</titleInfo>
<name type="personal">
<namePart type="given">Kevin</namePart>
<namePart type="family">Duh</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Helena</namePart>
<namePart type="family">Gomez</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Steven</namePart>
<namePart type="family">Bethard</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Mexico City, Mexico</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Ensuring the privacy of Large Language Models (LLMs) is becoming increasingly important. The most widely adopted technique to accomplish this is DP-SGD, which trains a model to guarantee Differential Privacy (DP). However, DP-SGD overestimates an adversary’s capabilities in having white box access to the model and, as a result, causes longer training times and larger memory usage than SGD. On the other hand, commercial LLM deployments are predominantly cloud-based; hence, adversarial access to LLMs is black-box. Motivated by these observations, we present Private Mixing of Ensemble Distributions (PMixED): a private prediction protocol for next-token prediction that utilizes the inherent stochasticity of next-token sampling and a public model to achieve Differential Privacy. We formalize this by introducing RD-mollifers which project each of the model’s output distribution from an ensemble of fine-tuned LLMs onto a set around a public LLM’s output distribution, then average the projected distributions and sample from it. Unlike DP-SGD which needs to consider the model architecture during training, PMixED is model agnostic, which makes PMixED a very appealing solution for current deployments. Our results show that PMixED achieves a stronger privacy guarantee than sample-level privacy and outperforms DP-SGD for privacy ε = 8 on large-scale datasets. Thus, PMixED offers a practical alternative to DP training methods for achieving strong generative utility without compromising privacy.</abstract>
<identifier type="citekey">flemings-etal-2024-differentially</identifier>
<location>
<url>https://aclanthology.org/2024.naacl-long.247</url>
</location>
<part>
<date>2024-06</date>
<extent unit="page">
<start>4390</start>
<end>4404</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Differentially Private Next-Token Prediction of Large Language Models
%A Flemings, James
%A Razaviyayn, Meisam
%A Annavaram, Murali
%Y Duh, Kevin
%Y Gomez, Helena
%Y Bethard, Steven
%S Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers)
%D 2024
%8 June
%I Association for Computational Linguistics
%C Mexico City, Mexico
%F flemings-etal-2024-differentially
%X Ensuring the privacy of Large Language Models (LLMs) is becoming increasingly important. The most widely adopted technique to accomplish this is DP-SGD, which trains a model to guarantee Differential Privacy (DP). However, DP-SGD overestimates an adversary’s capabilities in having white box access to the model and, as a result, causes longer training times and larger memory usage than SGD. On the other hand, commercial LLM deployments are predominantly cloud-based; hence, adversarial access to LLMs is black-box. Motivated by these observations, we present Private Mixing of Ensemble Distributions (PMixED): a private prediction protocol for next-token prediction that utilizes the inherent stochasticity of next-token sampling and a public model to achieve Differential Privacy. We formalize this by introducing RD-mollifers which project each of the model’s output distribution from an ensemble of fine-tuned LLMs onto a set around a public LLM’s output distribution, then average the projected distributions and sample from it. Unlike DP-SGD which needs to consider the model architecture during training, PMixED is model agnostic, which makes PMixED a very appealing solution for current deployments. Our results show that PMixED achieves a stronger privacy guarantee than sample-level privacy and outperforms DP-SGD for privacy ε = 8 on large-scale datasets. Thus, PMixED offers a practical alternative to DP training methods for achieving strong generative utility without compromising privacy.
%U https://aclanthology.org/2024.naacl-long.247
%P 4390-4404
Markdown (Informal)
[Differentially Private Next-Token Prediction of Large Language Models](https://aclanthology.org/2024.naacl-long.247) (Flemings et al., NAACL 2024)
ACL
- James Flemings, Meisam Razaviyayn, and Murali Annavaram. 2024. Differentially Private Next-Token Prediction of Large Language Models. In Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers), pages 4390–4404, Mexico City, Mexico. Association for Computational Linguistics.