A(DP)^2SGD: Asynchronous Decentralized Parallel Stochastic Gradient Descent with Differential Privacy

As deep learning models are usually massive and complex, distributed learning is essential for increasing training efficiency. Moreover, in many real-world application scenarios like healthcare, distributed learning can also keep the data local and protect privacy. Recently, the asynchronous decentr...

Ausführliche Beschreibung

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:IEEE transactions on pattern analysis and machine intelligence Jg. 44; H. 11; S. 1
Hauptverfasser: Xu, Jie, Zhang, Wei, Wang, Fei
Format: Journal Article
Sprache:Englisch
Veröffentlicht: 01.11.2022
ISSN:0162-8828, 1939-3539, 2160-9292, 1939-3539
Online-Zugang:Volltext
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
Abstract As deep learning models are usually massive and complex, distributed learning is essential for increasing training efficiency. Moreover, in many real-world application scenarios like healthcare, distributed learning can also keep the data local and protect privacy. Recently, the asynchronous decentralized parallel stochastic gradient descent (ADPSGD) algorithm has been proposed and demonstrated to be an efficient and practical strategy where there is no central server, so that each computing node only communicates with its neighbors. Although no raw data will be transmitted across different local nodes, there is still a risk of information leak during the communication process for malicious participants to make attacks. In this paper, we present a differentially private version of asynchronous decentralized parallel SGD framework, or A(DP) 2SGD for short, which maintains communication efficiency of ADPSGD and prevents the inference from malicious participants. Specifically, Rényi differential privacy is used to provide tighter privacy analysis for our composite Gaussian mechanisms while the convergence rate is consistent with the non-private version. Theoretical analysis shows A(DP) 2SGD also converges at the optimal O(1/√T) rate as SGD. Empirically, A(DP) 2SGD achieves comparable model accuracy as the differentially private version of Synchronous SGD (SSGD) but runs much faster than SSGD in heterogeneous computing environments.As deep learning models are usually massive and complex, distributed learning is essential for increasing training efficiency. Moreover, in many real-world application scenarios like healthcare, distributed learning can also keep the data local and protect privacy. Recently, the asynchronous decentralized parallel stochastic gradient descent (ADPSGD) algorithm has been proposed and demonstrated to be an efficient and practical strategy where there is no central server, so that each computing node only communicates with its neighbors. Although no raw data will be transmitted across different local nodes, there is still a risk of information leak during the communication process for malicious participants to make attacks. In this paper, we present a differentially private version of asynchronous decentralized parallel SGD framework, or A(DP) 2SGD for short, which maintains communication efficiency of ADPSGD and prevents the inference from malicious participants. Specifically, Rényi differential privacy is used to provide tighter privacy analysis for our composite Gaussian mechanisms while the convergence rate is consistent with the non-private version. Theoretical analysis shows A(DP) 2SGD also converges at the optimal O(1/√T) rate as SGD. Empirically, A(DP) 2SGD achieves comparable model accuracy as the differentially private version of Synchronous SGD (SSGD) but runs much faster than SSGD in heterogeneous computing environments.
AbstractList As deep learning models are usually massive and complex, distributed learning is essential for increasing training efficiency. Moreover, in many real-world application scenarios like healthcare, distributed learning can also keep the data local and protect privacy. Recently, the asynchronous decentralized parallel stochastic gradient descent (ADPSGD) algorithm has been proposed and demonstrated to be an efficient and practical strategy where there is no central server, so that each computing node only communicates with its neighbors. Although no raw data will be transmitted across different local nodes, there is still a risk of information leak during the communication process for malicious participants to make attacks. In this paper, we present a differentially private version of asynchronous decentralized parallel SGD framework, or A(DP) 2SGD for short, which maintains communication efficiency of ADPSGD and prevents the inference from malicious participants. Specifically, Rényi differential privacy is used to provide tighter privacy analysis for our composite Gaussian mechanisms while the convergence rate is consistent with the non-private version. Theoretical analysis shows A(DP) 2SGD also converges at the optimal O(1/√T) rate as SGD. Empirically, A(DP) 2SGD achieves comparable model accuracy as the differentially private version of Synchronous SGD (SSGD) but runs much faster than SSGD in heterogeneous computing environments.As deep learning models are usually massive and complex, distributed learning is essential for increasing training efficiency. Moreover, in many real-world application scenarios like healthcare, distributed learning can also keep the data local and protect privacy. Recently, the asynchronous decentralized parallel stochastic gradient descent (ADPSGD) algorithm has been proposed and demonstrated to be an efficient and practical strategy where there is no central server, so that each computing node only communicates with its neighbors. Although no raw data will be transmitted across different local nodes, there is still a risk of information leak during the communication process for malicious participants to make attacks. In this paper, we present a differentially private version of asynchronous decentralized parallel SGD framework, or A(DP) 2SGD for short, which maintains communication efficiency of ADPSGD and prevents the inference from malicious participants. Specifically, Rényi differential privacy is used to provide tighter privacy analysis for our composite Gaussian mechanisms while the convergence rate is consistent with the non-private version. Theoretical analysis shows A(DP) 2SGD also converges at the optimal O(1/√T) rate as SGD. Empirically, A(DP) 2SGD achieves comparable model accuracy as the differentially private version of Synchronous SGD (SSGD) but runs much faster than SSGD in heterogeneous computing environments.
Author Xu, Jie
Zhang, Wei
Wang, Fei
Author_xml – sequence: 1
  givenname: Jie
  surname: Xu
  fullname: Xu, Jie
– sequence: 2
  givenname: Wei
  surname: Zhang
  fullname: Zhang, Wei
– sequence: 3
  givenname: Fei
  surname: Wang
  fullname: Wang, Fei
BookMark eNp9kE9PwjAYhxuDiYB-AU874mHYP2u3eltAkQTjEjjbdF0XasqG7dDgp7cTTh489W3y_N68v2cEBk3baABuEZwiBPn9pshfllMMMZoSBNOUswswxIjBmGOOB2AIEcNxluHsCoy8f4cQJRSSIdD5ZF7cveH1Yv4Q5f7YqK1rm_bgo7lWuumctOZbV1Ehw2S1jdZdq7bSd0ZFCycrE5iA-p6Nvky3jeamrrULXyNtVDjzKdXxGlzW0np9c37HYPP0uJk9x6vXxXKWr2KFM9jFnPCUVjRRWSkzqnCdVhynrNSlSkpUIk4orZAmZaqx5iRUUgHGtU4ZY5KQMZic1u5d-3HQvhM7Ey6zVjY6VBKYMgYJThgNaHZClWu9d7oWynSyM21f2ViBoOjFil-xohcrzmJDFP-J7p3ZSXf8L_QDYuV9xA
CitedBy_id crossref_primary_10_1145_3650028
crossref_primary_10_1016_j_automatica_2024_111514
crossref_primary_10_1109_TPAMI_2023_3332428
crossref_primary_10_1155_2021_6679453
Cites_doi 10.1145/2810103.2813687
10.1109/CVPR.2018.00716
10.21437/Interspeech.2019-2700
10.1109/CSF.2017.11
10.1109/CVPR.2018.00474
10.1109/CVPR.2018.00745
10.1109/TNNLS.2019.2944481
10.1007/978-3-030-23502-4_10
10.1109/ICASSP.2019.8682888
10.1109/ICASSP40776.2020.9054065
10.29012/jpc.v4i1.612
10.1109/CVPR.2016.90
10.1109/CVPR.2017.634
10.4324/9781410605337-29
10.1007/s00287-019-01205-x
10.1007/11761679_29
10.1109/TII.2019.2942179
10.1145/2976749.2978318
10.29012/jpc.723
ContentType Journal Article
DBID AAYXX
CITATION
7X8
DOI 10.1109/TPAMI.2021.3107796
DatabaseName CrossRef
MEDLINE - Academic
DatabaseTitle CrossRef
MEDLINE - Academic
DatabaseTitleList MEDLINE - Academic
Database_xml – sequence: 1
  dbid: 7X8
  name: MEDLINE - Academic
  url: https://search.proquest.com/medline
  sourceTypes: Aggregation Database
DeliveryMethod fulltext_linktorsrc
Discipline Engineering
Computer Science
EISSN 2160-9292
1939-3539
EndPage 1
ExternalDocumentID 10_1109_TPAMI_2021_3107796
GroupedDBID ---
-DZ
-~X
.DC
0R~
29I
4.4
53G
5GY
5VS
6IK
97E
9M8
AAJGR
AASAJ
AAWTH
AAYXX
ABFSI
ABQJQ
ABVLG
ACGFO
ACGFS
ACIWK
ACNCT
ADRHT
AENEX
AETEA
AETIX
AGQYO
AGSQL
AHBIQ
AI.
AIBXA
AKJIK
AKQYR
ALLEH
ALMA_UNASSIGNED_HOLDINGS
ASUFR
ATWAV
BEFXN
BFFAM
BGNUA
BKEBE
BPEOZ
CITATION
CS3
DU5
E.L
EBS
EJD
F5P
FA8
HZ~
H~9
IBMZZ
ICLAB
IEDLZ
IFIPE
IFJZH
IPLJI
JAVBF
LAI
M43
MS~
O9-
OCL
P2P
PQQKQ
RIA
RIE
RNI
RNS
RXW
RZB
TAE
TN5
UHB
VH1
~02
7X8
ABAZT
ID FETCH-LOGICAL-c280t-93975d54c8ba85c2f7d9276bebc4b1b19355d1e3b7e2e93162c54c2fe7666a33
ISICitedReferencesCount 22
ISICitedReferencesURI http://www.webofscience.com/api/gateway?GWVersion=2&SrcApp=Summon&SrcAuth=ProQuest&DestLinkType=CitingArticles&DestApp=WOS_CPL&KeyUT=000864325900055&url=https%3A%2F%2Fcvtisr.summon.serialssolutions.com%2F%23%21%2Fsearch%3Fho%3Df%26include.ft.matches%3Dt%26l%3Dnull%26q%3D
ISSN 0162-8828
1939-3539
IngestDate Sun Sep 28 09:54:08 EDT 2025
Tue Nov 18 22:43:54 EST 2025
Sat Nov 29 05:16:01 EST 2025
IsPeerReviewed true
IsScholarly true
Issue 11
Language English
LinkModel OpenURL
MergedId FETCHMERGED-LOGICAL-c280t-93975d54c8ba85c2f7d9276bebc4b1b19355d1e3b7e2e93162c54c2fe7666a33
Notes ObjectType-Article-1
SourceType-Scholarly Journals-1
ObjectType-Feature-2
content type line 23
PQID 2566032465
PQPubID 23479
PageCount 1
ParticipantIDs proquest_miscellaneous_2566032465
crossref_citationtrail_10_1109_TPAMI_2021_3107796
crossref_primary_10_1109_TPAMI_2021_3107796
PublicationCentury 2000
PublicationDate 2022-11-01
PublicationDateYYYYMMDD 2022-11-01
PublicationDate_xml – month: 11
  year: 2022
  text: 2022-11-01
  day: 01
PublicationDecade 2020
PublicationTitle IEEE transactions on pattern analysis and machine intelligence
PublicationYear 2022
References ref13
Roy (ref7) 2019
ref12
ref15
Lian (ref11)
Cheng (ref20) 2018
ref36
ref31
Liu (ref34)
Zhang (ref2)
Li (ref25) 2019
ref32
Smith (ref3)
ref16
ref38
ref19
Bellet (ref22) 2017
ref18
Akiba (ref1)
Lalitha (ref9) 2019
Goyal (ref28) 2017
Bellet (ref23) 2017
Tan (ref35) 2019
ref24
ref45
ref26
ref42
ref41
ref44
ref21
ref43
Chen (ref29)
Lalitha (ref8) 2019
Chaudhuri (ref14) 2013; 14
Zhang (ref30)
Zinkevich (ref10)
Konevcny (ref4) 2016
ref5
Wang (ref27) 2019
ref40
McMahan (ref17) 2017
Shayan (ref6) 2018
Howard (ref37) 2017
Krizhevsky (ref33) 2009
Simonyan (ref39)
References_xml – ident: ref19
  doi: 10.1145/2810103.2813687
– ident: ref41
  doi: 10.1109/CVPR.2018.00716
– ident: ref45
  doi: 10.21437/Interspeech.2019-2700
– ident: ref39
  article-title: Very deep convolutional networks for large-scale image recognition
– year: 2017
  ident: ref28
  article-title: Accurate, large minibatch SGD: Training imagenet in 1 hour
– year: 2017
  ident: ref37
  article-title: Mobilenets: Efficient convolutional neural networks for mobile vision applications
– year: 2019
  ident: ref7
  article-title: Braintorrent: A peer-to-peer environment for decentralized federated learning
– start-page: 2350
  volume-title: Proc. 25th Int. Joint Conf. Artif. Intell.
  ident: ref2
  article-title: Staleness-aware async-SGD for distributed deep learning
– start-page: 4424
  volume-title: Proc. 31st Conf. Neural Inf. Process. Syst.
  ident: ref3
  article-title: Federated multi-task learning
– start-page: 3043
  volume-title: in Proc. 35th Int. Conf. Mach. Learn.
  ident: ref11
  article-title: Asynchronous decentralized parallel stochastic gradient descent
– ident: ref18
  doi: 10.1109/CSF.2017.11
– ident: ref38
  doi: 10.1109/CVPR.2018.00474
– ident: ref43
  doi: 10.1109/CVPR.2018.00745
– ident: ref5
  doi: 10.1109/TNNLS.2019.2944481
– ident: ref1
  article-title: ChainerMN: Scalable distributed deep learning framework
– year: 2018
  ident: ref6
  article-title: Biscotti: A ledger for private and secure peer-to-peer machine learning
– ident: ref21
  doi: 10.1007/978-3-030-23502-4_10
– ident: ref44
  doi: 10.1109/ICASSP.2019.8682888
– ident: ref31
  article-title: NCCL: Optimized primitives for collective multi-GPU communication.
– start-page: 171
  volume-title: Proc. IEEE Int. Conf. Data Mining
  ident: ref30
  article-title: Model accuracy and runtime tradeoff in distributed deep learning: A systematic study
– volume: 14
  start-page: 2905
  issue: 1
  year: 2013
  ident: ref14
  article-title: A near-optimal algorithm for differentially-private principal components
  publication-title: J. Mach. Learn. Res.
– year: 2019
  ident: ref25
  article-title: Asynchronous federated learning with differential privacy for edge intelligence
– year: 2016
  ident: ref4
  article-title: Federated learning: Strategies for improving communication efficiency
– ident: ref32
  doi: 10.1109/ICASSP40776.2020.9054065
– year: 2018
  ident: ref20
  article-title: LEASGD: An efficient and privacy-preserving decentralized algorithm for distributed learning
– ident: ref15
  doi: 10.29012/jpc.v4i1.612
– ident: ref29
  article-title: Revisiting distributed synchronous SGD
– ident: ref34
  article-title: Train CIFAR10 with PyTorch
– year: 2019
  ident: ref9
  article-title: Peer-to-peer federated learning on graphs
– year: 2019
  ident: ref27
  article-title: Efficient privacy-preserving nonconvex optimization
– ident: ref40
  doi: 10.1109/CVPR.2016.90
– ident: ref36
  doi: 10.1109/CVPR.2017.634
– year: 2017
  ident: ref17
  article-title: Learning differentially private recurrent language models
– year: 2017
  ident: ref22
  article-title: Fast and differentially private algorithms for decentralized collaborative machine learning
– ident: ref42
  doi: 10.4324/9781410605337-29
– year: 2009
  ident: ref33
  article-title: Learning multiple layers of features from tiny images
– year: 2017
  ident: ref23
  article-title: Personalized and private peer-to-peer machine learning
– ident: ref12
  doi: 10.1007/s00287-019-01205-x
– ident: ref13
  doi: 10.1007/11761679_29
– year: 2019
  ident: ref8
  article-title: Decentralized Bayesian learning over graphs
– ident: ref24
  doi: 10.1109/TII.2019.2942179
– year: 2019
  ident: ref35
  article-title: EfficientNet: Rethinking model scaling for convolutional neural networks
– ident: ref16
  doi: 10.1145/2976749.2978318
– start-page: 2595
  volume-title: Proc. 24th Annu. Conf. Neural Inf. Process. Syst.
  ident: ref10
  article-title: Parallelized stochastic gradient descent
– ident: ref26
  doi: 10.29012/jpc.723
SSID ssj0014503
Score 2.5588422
Snippet As deep learning models are usually massive and complex, distributed learning is essential for increasing training efficiency. Moreover, in many real-world...
SourceID proquest
crossref
SourceType Aggregation Database
Enrichment Source
Index Database
StartPage 1
Title A(DP)^2SGD: Asynchronous Decentralized Parallel Stochastic Gradient Descent with Differential Privacy
URI https://www.proquest.com/docview/2566032465
Volume 44
WOSCitedRecordID wos000864325900055&url=https%3A%2F%2Fcvtisr.summon.serialssolutions.com%2F%23%21%2Fsearch%3Fho%3Df%26include.ft.matches%3Dt%26l%3Dnull%26q%3D
hasFullText 1
inHoldings 1
isFullTextHit
isPrint
journalDatabaseRights – providerCode: PRVIEE
  databaseName: IEEE Electronic Library (IEL)
  customDbUrl:
  eissn: 2160-9292
  dateEnd: 99991231
  omitProxy: false
  ssIdentifier: ssj0014503
  issn: 0162-8828
  databaseCode: RIE
  dateStart: 19790101
  isFulltext: true
  titleUrlDefault: https://ieeexplore.ieee.org/
  providerName: IEEE
link http://cvtisr.summon.serialssolutions.com/2.0.0/link/0/eLvHCXMwtV1Lb9QwELa2hQM98CiglkdlJJBAVUrsxLHDbUXawoEqEiuxJ6LYcUSkkq7SdFX4I_xdxo6TzUJB9MAlWo0ca-P5NJ4Zj79B6LkUsQq4L7yyFIUHBo94AvY1r9Cw94WUwFvKNpvgJydiPo_TyeRHfxdmecrrWlxexov_qmqQgbLN1dlrqHuYFATwG5QOT1A7PP9J8VPwGZMUIvwX7JB-PE5s6u_8W60MDa4peE20q8isvoO3meaNaadizqfP1Jfc0DbvHze2EKyFoZbsqcvWJq6XSmuS7GlTLXO1diZs4kbTcqLvP24PIhaWv9NUPDvyE5Oo_2orOLXlquj5QHvNzy8ssKpBMKS0P-lqlf7vREdO5NIWEPGSIW3hMpkRmGLhboZrK6Mk8j3w1-jIopLR1kyutvmWMnWWTj-8h3ifkgPwWDmPryDY_mXjG8oRbSDkx5mdIzNzZG6ODXSDchaL7mLgcDwVMttye_iG_jaWH7_-_X-sezzrG771YmZ30W0XfuBpB5t7aKLrbXSnb-2BnaXfRlsjnsr7SE9fJumrzwZNb_AYS3gNS7jHEl5hCfdYwg5L2GAJj7GEHZYeoNnR4eztO8_15_AUFX7rxeDLsoKFSshcMEVLXsSUR1JLFUoiiaHuL4gOJNdUxwEsloLBtNQcYuY8CB6izfqs1jsIizwC01DIgkdlyGkhWSGlH4J1YSVELGwXkX4JM-W4600LldPsz6rbRfvDO4uOueWvo5_1msnAwJpTs7zWsJAZxASRD2FHxB5da8bH6NYK90_QZttc6Kfoplq21Xmzhzb4XOxZVP0Es7eaoQ
linkProvider IEEE
openUrl ctx_ver=Z39.88-2004&ctx_enc=info%3Aofi%2Fenc%3AUTF-8&rfr_id=info%3Asid%2Fsummon.serialssolutions.com&rft_val_fmt=info%3Aofi%2Ffmt%3Akev%3Amtx%3Ajournal&rft.genre=article&rft.atitle=A%28DP%29%5E2SGD%3A+Asynchronous+Decentralized+Parallel+Stochastic+Gradient+Descent+with+Differential+Privacy&rft.jtitle=IEEE+transactions+on+pattern+analysis+and+machine+intelligence&rft.au=Xu%2C+Jie&rft.au=Zhang%2C+Wei&rft.au=Wang%2C+Fei&rft.date=2022-11-01&rft.issn=0162-8828&rft.eissn=2160-9292&rft.spage=1&rft.epage=1&rft_id=info:doi/10.1109%2FTPAMI.2021.3107796&rft.externalDBID=n%2Fa&rft.externalDocID=10_1109_TPAMI_2021_3107796
thumbnail_l http://covers-cdn.summon.serialssolutions.com/index.aspx?isbn=/lc.gif&issn=0162-8828&client=summon
thumbnail_m http://covers-cdn.summon.serialssolutions.com/index.aspx?isbn=/mc.gif&issn=0162-8828&client=summon
thumbnail_s http://covers-cdn.summon.serialssolutions.com/index.aspx?isbn=/sc.gif&issn=0162-8828&client=summon