Reconstructing training data from document understanding models

Document understanding models are increasingly employed by companies to supplant humans in processing sensitive documents, such as invoices, tax notices, or even ID cards. However, the robustness of such models to privacy attacks remains vastly unexplored. This paper presents CDMI, the first reconst...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:arXiv.org 2024-06
Hauptverfasser: Dentan, Jérémie, Paran, Arnaud, Shabou, Aymen
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title arXiv.org
container_volume
creator Dentan, Jérémie
Paran, Arnaud
Shabou, Aymen
description Document understanding models are increasingly employed by companies to supplant humans in processing sensitive documents, such as invoices, tax notices, or even ID cards. However, the robustness of such models to privacy attacks remains vastly unexplored. This paper presents CDMI, the first reconstruction attack designed to extract sensitive fields from the training data of these models. We attack LayoutLM and BROS architectures, demonstrating that an adversary can perfectly reconstruct up to 4.1% of the fields of the documents used for fine-tuning, including some names, dates, and invoice amounts up to six-digit numbers. When our reconstruction attack is combined with a membership inference attack, our attack accuracy escalates to 22.5%. In addition, we introduce two new end-to-end metrics and evaluate our approach under various conditions: unimodal or bimodal data, LayoutLM or BROS backbones, four fine-tuning tasks, and two public datasets (FUNSD and SROIE). We also investigate the interplay between overfitting, predictive performance, and susceptibility to our attack. We conclude with a discussion on possible defenses against our attack and potential future research directions to construct robust document understanding models.
format Article
fullrecord <record><control><sourceid>proquest</sourceid><recordid>TN_cdi_proquest_journals_3065127294</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>3065127294</sourcerecordid><originalsourceid>FETCH-proquest_journals_30651272943</originalsourceid><addsrcrecordid>eNqNikEKwjAQAIMgWLR_CHgupJum1ZMHUTyL9xKaVFrajWY3_9eCD_A0AzMrkYHWZXGoADYiJxqVUlA3YIzOxOnuu4DEMXU84FNytAMu4ixb2ccwSxe6NHtkmdD5SGzRLcMcnJ9oJ9a9ncjnP27F_np5nG_FK4Z38sTtGFLEb2q1qk0JDRwr_d_1AXp4Oa8</addsrcrecordid><sourcetype>Aggregation Database</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype><pqid>3065127294</pqid></control><display><type>article</type><title>Reconstructing training data from document understanding models</title><source>Free E- Journals</source><creator>Dentan, Jérémie ; Paran, Arnaud ; Shabou, Aymen</creator><creatorcontrib>Dentan, Jérémie ; Paran, Arnaud ; Shabou, Aymen</creatorcontrib><description>Document understanding models are increasingly employed by companies to supplant humans in processing sensitive documents, such as invoices, tax notices, or even ID cards. However, the robustness of such models to privacy attacks remains vastly unexplored. This paper presents CDMI, the first reconstruction attack designed to extract sensitive fields from the training data of these models. We attack LayoutLM and BROS architectures, demonstrating that an adversary can perfectly reconstruct up to 4.1% of the fields of the documents used for fine-tuning, including some names, dates, and invoice amounts up to six-digit numbers. When our reconstruction attack is combined with a membership inference attack, our attack accuracy escalates to 22.5%. In addition, we introduce two new end-to-end metrics and evaluate our approach under various conditions: unimodal or bimodal data, LayoutLM or BROS backbones, four fine-tuning tasks, and two public datasets (FUNSD and SROIE). We also investigate the interplay between overfitting, predictive performance, and susceptibility to our attack. We conclude with a discussion on possible defenses against our attack and potential future research directions to construct robust document understanding models.</description><identifier>EISSN: 2331-8422</identifier><language>eng</language><publisher>Ithaca: Cornell University Library, arXiv.org</publisher><subject>Documents ; Performance prediction ; Reconstruction</subject><ispartof>arXiv.org, 2024-06</ispartof><rights>2024. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>780,784</link.rule.ids></links><search><creatorcontrib>Dentan, Jérémie</creatorcontrib><creatorcontrib>Paran, Arnaud</creatorcontrib><creatorcontrib>Shabou, Aymen</creatorcontrib><title>Reconstructing training data from document understanding models</title><title>arXiv.org</title><description>Document understanding models are increasingly employed by companies to supplant humans in processing sensitive documents, such as invoices, tax notices, or even ID cards. However, the robustness of such models to privacy attacks remains vastly unexplored. This paper presents CDMI, the first reconstruction attack designed to extract sensitive fields from the training data of these models. We attack LayoutLM and BROS architectures, demonstrating that an adversary can perfectly reconstruct up to 4.1% of the fields of the documents used for fine-tuning, including some names, dates, and invoice amounts up to six-digit numbers. When our reconstruction attack is combined with a membership inference attack, our attack accuracy escalates to 22.5%. In addition, we introduce two new end-to-end metrics and evaluate our approach under various conditions: unimodal or bimodal data, LayoutLM or BROS backbones, four fine-tuning tasks, and two public datasets (FUNSD and SROIE). We also investigate the interplay between overfitting, predictive performance, and susceptibility to our attack. We conclude with a discussion on possible defenses against our attack and potential future research directions to construct robust document understanding models.</description><subject>Documents</subject><subject>Performance prediction</subject><subject>Reconstruction</subject><issn>2331-8422</issn><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2024</creationdate><recordtype>article</recordtype><sourceid>ABUWG</sourceid><sourceid>AFKRA</sourceid><sourceid>AZQEC</sourceid><sourceid>BENPR</sourceid><sourceid>CCPQU</sourceid><sourceid>DWQXO</sourceid><recordid>eNqNikEKwjAQAIMgWLR_CHgupJum1ZMHUTyL9xKaVFrajWY3_9eCD_A0AzMrkYHWZXGoADYiJxqVUlA3YIzOxOnuu4DEMXU84FNytAMu4ixb2ccwSxe6NHtkmdD5SGzRLcMcnJ9oJ9a9ncjnP27F_np5nG_FK4Z38sTtGFLEb2q1qk0JDRwr_d_1AXp4Oa8</recordid><startdate>20240605</startdate><enddate>20240605</enddate><creator>Dentan, Jérémie</creator><creator>Paran, Arnaud</creator><creator>Shabou, Aymen</creator><general>Cornell University Library, arXiv.org</general><scope>8FE</scope><scope>8FG</scope><scope>ABJCF</scope><scope>ABUWG</scope><scope>AFKRA</scope><scope>AZQEC</scope><scope>BENPR</scope><scope>BGLVJ</scope><scope>CCPQU</scope><scope>DWQXO</scope><scope>HCIFZ</scope><scope>L6V</scope><scope>M7S</scope><scope>PIMPY</scope><scope>PQEST</scope><scope>PQQKQ</scope><scope>PQUKI</scope><scope>PRINS</scope><scope>PTHSS</scope></search><sort><creationdate>20240605</creationdate><title>Reconstructing training data from document understanding models</title><author>Dentan, Jérémie ; Paran, Arnaud ; Shabou, Aymen</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-proquest_journals_30651272943</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2024</creationdate><topic>Documents</topic><topic>Performance prediction</topic><topic>Reconstruction</topic><toplevel>online_resources</toplevel><creatorcontrib>Dentan, Jérémie</creatorcontrib><creatorcontrib>Paran, Arnaud</creatorcontrib><creatorcontrib>Shabou, Aymen</creatorcontrib><collection>ProQuest SciTech Collection</collection><collection>ProQuest Technology Collection</collection><collection>Materials Science &amp; Engineering Collection</collection><collection>ProQuest Central (Alumni Edition)</collection><collection>ProQuest Central UK/Ireland</collection><collection>ProQuest Central Essentials</collection><collection>ProQuest Central</collection><collection>Technology Collection</collection><collection>ProQuest One Community College</collection><collection>ProQuest Central Korea</collection><collection>SciTech Premium Collection</collection><collection>ProQuest Engineering Collection</collection><collection>Engineering Database</collection><collection>Publicly Available Content Database</collection><collection>ProQuest One Academic Eastern Edition (DO NOT USE)</collection><collection>ProQuest One Academic</collection><collection>ProQuest One Academic UKI Edition</collection><collection>ProQuest Central China</collection><collection>Engineering Collection</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext</fulltext></delivery><addata><au>Dentan, Jérémie</au><au>Paran, Arnaud</au><au>Shabou, Aymen</au><format>book</format><genre>document</genre><ristype>GEN</ristype><atitle>Reconstructing training data from document understanding models</atitle><jtitle>arXiv.org</jtitle><date>2024-06-05</date><risdate>2024</risdate><eissn>2331-8422</eissn><abstract>Document understanding models are increasingly employed by companies to supplant humans in processing sensitive documents, such as invoices, tax notices, or even ID cards. However, the robustness of such models to privacy attacks remains vastly unexplored. This paper presents CDMI, the first reconstruction attack designed to extract sensitive fields from the training data of these models. We attack LayoutLM and BROS architectures, demonstrating that an adversary can perfectly reconstruct up to 4.1% of the fields of the documents used for fine-tuning, including some names, dates, and invoice amounts up to six-digit numbers. When our reconstruction attack is combined with a membership inference attack, our attack accuracy escalates to 22.5%. In addition, we introduce two new end-to-end metrics and evaluate our approach under various conditions: unimodal or bimodal data, LayoutLM or BROS backbones, four fine-tuning tasks, and two public datasets (FUNSD and SROIE). We also investigate the interplay between overfitting, predictive performance, and susceptibility to our attack. We conclude with a discussion on possible defenses against our attack and potential future research directions to construct robust document understanding models.</abstract><cop>Ithaca</cop><pub>Cornell University Library, arXiv.org</pub><oa>free_for_read</oa></addata></record>
fulltext fulltext
identifier EISSN: 2331-8422
ispartof arXiv.org, 2024-06
issn 2331-8422
language eng
recordid cdi_proquest_journals_3065127294
source Free E- Journals
subjects Documents
Performance prediction
Reconstruction
title Reconstructing training data from document understanding models
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-01-04T09%3A49%3A35IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-proquest&rft_val_fmt=info:ofi/fmt:kev:mtx:book&rft.genre=document&rft.atitle=Reconstructing%20training%20data%20from%20document%20understanding%20models&rft.jtitle=arXiv.org&rft.au=Dentan,%20J%C3%A9r%C3%A9mie&rft.date=2024-06-05&rft.eissn=2331-8422&rft_id=info:doi/&rft_dat=%3Cproquest%3E3065127294%3C/proquest%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_pqid=3065127294&rft_id=info:pmid/&rfr_iscdi=true