On the Convergence and Calibration of Deep Learning with Differential Privacy

Differentially private (DP) training preserves the data privacy usually at the cost of slower convergence (and thus lower accuracy), as well as more severe mis-calibration than its non-private counterpart. To analyze the convergence of DP training, we formulate a continuous time analysis through the...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Bu, Zhiqi, Wang, Hua, Dai, Zongyu, Long, Qi
Format: Artikel
Sprache:eng
Schlagworte:
Online-Zugang:Volltext bestellen
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
container_end_page
container_issue
container_start_page
container_title
container_volume
creator Bu, Zhiqi
Wang, Hua
Dai, Zongyu
Long, Qi
description Differentially private (DP) training preserves the data privacy usually at the cost of slower convergence (and thus lower accuracy), as well as more severe mis-calibration than its non-private counterpart. To analyze the convergence of DP training, we formulate a continuous time analysis through the lens of neural tangent kernel (NTK), which characterizes the per-sample gradient clipping and the noise addition in DP training, for arbitrary network architectures and loss functions. Interestingly, we show that the noise addition only affects the privacy risk but not the convergence or calibration, whereas the per-sample gradient clipping (under both flat and layerwise clipping styles) only affects the convergence and calibration. Furthermore, we observe that while DP models trained with small clipping norm usually achieve the best accurate, but are poorly calibrated and thus unreliable. In sharp contrast, DP models trained with large clipping norm enjoy the same privacy guarantee and similar accuracy, but are significantly more \textit{calibrated}. Our code can be found at \url{https://github.com/woodyx218/opacus_global_clipping}.
doi_str_mv 10.48550/arxiv.2106.07830
format Article
fullrecord <record><control><sourceid>arxiv_GOX</sourceid><recordid>TN_cdi_arxiv_primary_2106_07830</recordid><sourceformat>XML</sourceformat><sourcesystem>PC</sourcesystem><sourcerecordid>2106_07830</sourcerecordid><originalsourceid>FETCH-LOGICAL-a670-2caed5191cff3dac614450803749a6e0e84f31ab3cae87dfa9ce965286b3e5b93</originalsourceid><addsrcrecordid>eNotz71OwzAUhmEvDKjlApg4N5Bgxz-xxyrlTwoqQ_foxDluLQWnMlGgdw8Upm959UkPY7eCl8pqze8xf8WlrAQ3Ja-t5NfsdZdgPhI0U1ooHyh5AkwDNDjGPuMcpwRTgC3RCVrCnGI6wGecj7CNIVCmNEcc4S3HBf15za4Cjh90878rtn982DfPRbt7emk2bYGm5kXlkQYtnPAhyAG9EUppbrmslUNDnKwKUmAvfzpbDwGdJ2d0ZU0vSfdOrtjd3-3F051yfMd87n5d3cUlvwEKc0iR</addsrcrecordid><sourcetype>Open Access Repository</sourcetype><iscdi>true</iscdi><recordtype>article</recordtype></control><display><type>article</type><title>On the Convergence and Calibration of Deep Learning with Differential Privacy</title><source>arXiv.org</source><creator>Bu, Zhiqi ; Wang, Hua ; Dai, Zongyu ; Long, Qi</creator><creatorcontrib>Bu, Zhiqi ; Wang, Hua ; Dai, Zongyu ; Long, Qi</creatorcontrib><description>Differentially private (DP) training preserves the data privacy usually at the cost of slower convergence (and thus lower accuracy), as well as more severe mis-calibration than its non-private counterpart. To analyze the convergence of DP training, we formulate a continuous time analysis through the lens of neural tangent kernel (NTK), which characterizes the per-sample gradient clipping and the noise addition in DP training, for arbitrary network architectures and loss functions. Interestingly, we show that the noise addition only affects the privacy risk but not the convergence or calibration, whereas the per-sample gradient clipping (under both flat and layerwise clipping styles) only affects the convergence and calibration. Furthermore, we observe that while DP models trained with small clipping norm usually achieve the best accurate, but are poorly calibrated and thus unreliable. In sharp contrast, DP models trained with large clipping norm enjoy the same privacy guarantee and similar accuracy, but are significantly more \textit{calibrated}. Our code can be found at \url{https://github.com/woodyx218/opacus_global_clipping}.</description><identifier>DOI: 10.48550/arxiv.2106.07830</identifier><language>eng</language><subject>Computer Science - Learning ; Statistics - Machine Learning</subject><creationdate>2021-06</creationdate><rights>http://arxiv.org/licenses/nonexclusive-distrib/1.0</rights><oa>free_for_read</oa><woscitedreferencessubscribed>false</woscitedreferencessubscribed></display><links><openurl>$$Topenurl_article</openurl><openurlfulltext>$$Topenurlfull_article</openurlfulltext><thumbnail>$$Tsyndetics_thumb_exl</thumbnail><link.rule.ids>228,230,776,881</link.rule.ids><linktorsrc>$$Uhttps://arxiv.org/abs/2106.07830$$EView_record_in_Cornell_University$$FView_record_in_$$GCornell_University$$Hfree_for_read</linktorsrc><backlink>$$Uhttps://doi.org/10.48550/arXiv.2106.07830$$DView paper in arXiv$$Hfree_for_read</backlink></links><search><creatorcontrib>Bu, Zhiqi</creatorcontrib><creatorcontrib>Wang, Hua</creatorcontrib><creatorcontrib>Dai, Zongyu</creatorcontrib><creatorcontrib>Long, Qi</creatorcontrib><title>On the Convergence and Calibration of Deep Learning with Differential Privacy</title><description>Differentially private (DP) training preserves the data privacy usually at the cost of slower convergence (and thus lower accuracy), as well as more severe mis-calibration than its non-private counterpart. To analyze the convergence of DP training, we formulate a continuous time analysis through the lens of neural tangent kernel (NTK), which characterizes the per-sample gradient clipping and the noise addition in DP training, for arbitrary network architectures and loss functions. Interestingly, we show that the noise addition only affects the privacy risk but not the convergence or calibration, whereas the per-sample gradient clipping (under both flat and layerwise clipping styles) only affects the convergence and calibration. Furthermore, we observe that while DP models trained with small clipping norm usually achieve the best accurate, but are poorly calibrated and thus unreliable. In sharp contrast, DP models trained with large clipping norm enjoy the same privacy guarantee and similar accuracy, but are significantly more \textit{calibrated}. Our code can be found at \url{https://github.com/woodyx218/opacus_global_clipping}.</description><subject>Computer Science - Learning</subject><subject>Statistics - Machine Learning</subject><fulltext>true</fulltext><rsrctype>article</rsrctype><creationdate>2021</creationdate><recordtype>article</recordtype><sourceid>GOX</sourceid><recordid>eNotz71OwzAUhmEvDKjlApg4N5Bgxz-xxyrlTwoqQ_foxDluLQWnMlGgdw8Upm959UkPY7eCl8pqze8xf8WlrAQ3Ja-t5NfsdZdgPhI0U1ooHyh5AkwDNDjGPuMcpwRTgC3RCVrCnGI6wGecj7CNIVCmNEcc4S3HBf15za4Cjh90878rtn982DfPRbt7emk2bYGm5kXlkQYtnPAhyAG9EUppbrmslUNDnKwKUmAvfzpbDwGdJ2d0ZU0vSfdOrtjd3-3F051yfMd87n5d3cUlvwEKc0iR</recordid><startdate>20210614</startdate><enddate>20210614</enddate><creator>Bu, Zhiqi</creator><creator>Wang, Hua</creator><creator>Dai, Zongyu</creator><creator>Long, Qi</creator><scope>AKY</scope><scope>EPD</scope><scope>GOX</scope></search><sort><creationdate>20210614</creationdate><title>On the Convergence and Calibration of Deep Learning with Differential Privacy</title><author>Bu, Zhiqi ; Wang, Hua ; Dai, Zongyu ; Long, Qi</author></sort><facets><frbrtype>5</frbrtype><frbrgroupid>cdi_FETCH-LOGICAL-a670-2caed5191cff3dac614450803749a6e0e84f31ab3cae87dfa9ce965286b3e5b93</frbrgroupid><rsrctype>articles</rsrctype><prefilter>articles</prefilter><language>eng</language><creationdate>2021</creationdate><topic>Computer Science - Learning</topic><topic>Statistics - Machine Learning</topic><toplevel>online_resources</toplevel><creatorcontrib>Bu, Zhiqi</creatorcontrib><creatorcontrib>Wang, Hua</creatorcontrib><creatorcontrib>Dai, Zongyu</creatorcontrib><creatorcontrib>Long, Qi</creatorcontrib><collection>arXiv Computer Science</collection><collection>arXiv Statistics</collection><collection>arXiv.org</collection></facets><delivery><delcategory>Remote Search Resource</delcategory><fulltext>fulltext_linktorsrc</fulltext></delivery><addata><au>Bu, Zhiqi</au><au>Wang, Hua</au><au>Dai, Zongyu</au><au>Long, Qi</au><format>journal</format><genre>article</genre><ristype>JOUR</ristype><atitle>On the Convergence and Calibration of Deep Learning with Differential Privacy</atitle><date>2021-06-14</date><risdate>2021</risdate><abstract>Differentially private (DP) training preserves the data privacy usually at the cost of slower convergence (and thus lower accuracy), as well as more severe mis-calibration than its non-private counterpart. To analyze the convergence of DP training, we formulate a continuous time analysis through the lens of neural tangent kernel (NTK), which characterizes the per-sample gradient clipping and the noise addition in DP training, for arbitrary network architectures and loss functions. Interestingly, we show that the noise addition only affects the privacy risk but not the convergence or calibration, whereas the per-sample gradient clipping (under both flat and layerwise clipping styles) only affects the convergence and calibration. Furthermore, we observe that while DP models trained with small clipping norm usually achieve the best accurate, but are poorly calibrated and thus unreliable. In sharp contrast, DP models trained with large clipping norm enjoy the same privacy guarantee and similar accuracy, but are significantly more \textit{calibrated}. Our code can be found at \url{https://github.com/woodyx218/opacus_global_clipping}.</abstract><doi>10.48550/arxiv.2106.07830</doi><oa>free_for_read</oa></addata></record>
fulltext fulltext_linktorsrc
identifier DOI: 10.48550/arxiv.2106.07830
ispartof
issn
language eng
recordid cdi_arxiv_primary_2106_07830
source arXiv.org
subjects Computer Science - Learning
Statistics - Machine Learning
title On the Convergence and Calibration of Deep Learning with Differential Privacy
url https://sfx.bib-bvb.de/sfx_tum?ctx_ver=Z39.88-2004&ctx_enc=info:ofi/enc:UTF-8&ctx_tim=2025-02-08T04%3A26%3A40IST&url_ver=Z39.88-2004&url_ctx_fmt=infofi/fmt:kev:mtx:ctx&rfr_id=info:sid/primo.exlibrisgroup.com:primo3-Article-arxiv_GOX&rft_val_fmt=info:ofi/fmt:kev:mtx:journal&rft.genre=article&rft.atitle=On%20the%20Convergence%20and%20Calibration%20of%20Deep%20Learning%20with%20Differential%20Privacy&rft.au=Bu,%20Zhiqi&rft.date=2021-06-14&rft_id=info:doi/10.48550/arxiv.2106.07830&rft_dat=%3Carxiv_GOX%3E2106_07830%3C/arxiv_GOX%3E%3Curl%3E%3C/url%3E&disable_directlink=true&sfx.directlink=off&sfx.report_link=0&rft_id=info:oai/&rft_id=info:pmid/&rfr_iscdi=true