@inproceedings{ippolito-etal-2023-reverse,
title = "Reverse-Engineering Decoding Strategies Given Blackbox Access to a Language Generation System",
author = "Ippolito, Daphne and
Carlini, Nicholas and
Lee, Katherine and
Nasr, Milad and
Yu, Yun William",
editor = "Keet, C. Maria and
Lee, Hung-Yi and
Zarrie{\ss}, Sina",
booktitle = "Proceedings of the 16th International Natural Language Generation Conference",
month = sep,
year = "2023",
address = "Prague, Czechia",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2023.inlg-main.28",
doi = "10.18653/v1/2023.inlg-main.28",
pages = "396--406",
abstract = "Neural language models are increasingly deployed into APIs and websites that allow a user to pass in a prompt and receive generated text. Many of these systems do not reveal generation parameters. In this paper, we present methods to reverse-engineer the decoding method used to generate text (i.e., top-{\_}k{\_} or nucleus sampling). Our ability to discover which decoding strategy was used has implications for detecting generated text. Additionally, the process of discovering the decoding strategy can reveal biases caused by selecting decoding settings which severely truncate a model{'}s predicted distributions. We perform our attack on several families of open-source language models, as well as on production systems (e.g., ChatGPT).",
}
<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="ippolito-etal-2023-reverse">
<titleInfo>
<title>Reverse-Engineering Decoding Strategies Given Blackbox Access to a Language Generation System</title>
</titleInfo>
<name type="personal">
<namePart type="given">Daphne</namePart>
<namePart type="family">Ippolito</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nicholas</namePart>
<namePart type="family">Carlini</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Katherine</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Milad</namePart>
<namePart type="family">Nasr</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yun</namePart>
<namePart type="given">William</namePart>
<namePart type="family">Yu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2023-09</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 16th International Natural Language Generation Conference</title>
</titleInfo>
<name type="personal">
<namePart type="given">C</namePart>
<namePart type="given">Maria</namePart>
<namePart type="family">Keet</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sina</namePart>
<namePart type="family">Zarrieß</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Prague, Czechia</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Neural language models are increasingly deployed into APIs and websites that allow a user to pass in a prompt and receive generated text. Many of these systems do not reveal generation parameters. In this paper, we present methods to reverse-engineer the decoding method used to generate text (i.e., top-_k_ or nucleus sampling). Our ability to discover which decoding strategy was used has implications for detecting generated text. Additionally, the process of discovering the decoding strategy can reveal biases caused by selecting decoding settings which severely truncate a model’s predicted distributions. We perform our attack on several families of open-source language models, as well as on production systems (e.g., ChatGPT).</abstract>
<identifier type="citekey">ippolito-etal-2023-reverse</identifier>
<identifier type="doi">10.18653/v1/2023.inlg-main.28</identifier>
<location>
<url>https://aclanthology.org/2023.inlg-main.28</url>
</location>
<part>
<date>2023-09</date>
<extent unit="page">
<start>396</start>
<end>406</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Reverse-Engineering Decoding Strategies Given Blackbox Access to a Language Generation System
%A Ippolito, Daphne
%A Carlini, Nicholas
%A Lee, Katherine
%A Nasr, Milad
%A Yu, Yun William
%Y Keet, C. Maria
%Y Lee, Hung-Yi
%Y Zarrieß, Sina
%S Proceedings of the 16th International Natural Language Generation Conference
%D 2023
%8 September
%I Association for Computational Linguistics
%C Prague, Czechia
%F ippolito-etal-2023-reverse
%X Neural language models are increasingly deployed into APIs and websites that allow a user to pass in a prompt and receive generated text. Many of these systems do not reveal generation parameters. In this paper, we present methods to reverse-engineer the decoding method used to generate text (i.e., top-_k_ or nucleus sampling). Our ability to discover which decoding strategy was used has implications for detecting generated text. Additionally, the process of discovering the decoding strategy can reveal biases caused by selecting decoding settings which severely truncate a model’s predicted distributions. We perform our attack on several families of open-source language models, as well as on production systems (e.g., ChatGPT).
%R 10.18653/v1/2023.inlg-main.28
%U https://aclanthology.org/2023.inlg-main.28
%U https://doi.org/10.18653/v1/2023.inlg-main.28
%P 396-406
Markdown (Informal)
[Reverse-Engineering Decoding Strategies Given Blackbox Access to a Language Generation System](https://aclanthology.org/2023.inlg-main.28) (Ippolito et al., INLG-SIGDIAL 2023)
ACL