An Investigation on Sparsity of CapsNets for Adversarial Robustness

Lei Zhao; Lei Huang

doi:10.1145/3475724.3483609

An Investigation on Sparsity of CapsNets for Adversarial Robustness

Lei Zhao
, Lei Huang^*

^*Corresponding author for this work

School of Artificial Intelligence(Institute of Artificial Intelligence)

DiDi Chuxing

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

2 Scopus citations

Abstract

The routing-by-agreement mechanism in capsule networks (CapsNets) is used to build visual hierarchical relationships with a characteristic of assigning parts to wholes. The connections between capsules of different layers become sparser with more iterations of routing. This paper proposes techniques in measuring, controlling, and visualizing the sparsity of CapsNets. One essential observation in this paper is that the sparser CapsNets are possibly more robust to the adversarial attacks. We believe this observation will provide insights into designing more robust models.

Original language	English
Title of host publication	AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021
Publisher	Association for Computing Machinery, Inc
Pages	55-61
Number of pages	7
ISBN (Electronic)	9781450386722
DOIs	https://doi.org/10.1145/3475724.3483609
State	Published - 22 Oct 2021
Event	1st International Workshop on Adversarial Learning for Multimedia, AdvM 2021, co-located with ACM MM 2021 - Virtual, Online, China Duration: 20 Oct 2021 → 20 Oct 2021

Publication series

Name	AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021

Conference

Conference	1st International Workshop on Adversarial Learning for Multimedia, AdvM 2021, co-located with ACM MM 2021
Country/Territory	China
City	Virtual, Online
Period	20/10/21 → 20/10/21

Keywords

adversarial robustness
capsnet
class activation maps
sparsity

Access to Document

10.1145/3475724.3483609

Cite this

Zhao, L., & Huang, L. (2021). An Investigation on Sparsity of CapsNets for Adversarial Robustness. In AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021 (pp. 55-61). (AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021). Association for Computing Machinery, Inc. https://doi.org/10.1145/3475724.3483609

Zhao, Lei ; Huang, Lei. / An Investigation on Sparsity of CapsNets for Adversarial Robustness. AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021. Association for Computing Machinery, Inc, 2021. pp. 55-61 (AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021).

@inproceedings{e4da80bdb09f4f1e9180d121a11f848f,

title = "An Investigation on Sparsity of CapsNets for Adversarial Robustness",

abstract = "The routing-by-agreement mechanism in capsule networks (CapsNets) is used to build visual hierarchical relationships with a characteristic of assigning parts to wholes. The connections between capsules of different layers become sparser with more iterations of routing. This paper proposes techniques in measuring, controlling, and visualizing the sparsity of CapsNets. One essential observation in this paper is that the sparser CapsNets are possibly more robust to the adversarial attacks. We believe this observation will provide insights into designing more robust models.",

keywords = "adversarial robustness, capsnet, class activation maps, sparsity",

author = "Lei Zhao and Lei Huang",

note = "Publisher Copyright: {\textcopyright} 2021 ACM.; 1st International Workshop on Adversarial Learning for Multimedia, AdvM 2021, co-located with ACM MM 2021 ; Conference date: 20-10-2021 Through 20-10-2021",

year = "2021",

month = oct,

day = "22",

doi = "10.1145/3475724.3483609",

language = "英语",

series = "AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021",

publisher = "Association for Computing Machinery, Inc",

pages = "55--61",

booktitle = "AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021",

}

Zhao, L & Huang, L 2021, An Investigation on Sparsity of CapsNets for Adversarial Robustness. in AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021. AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021, Association for Computing Machinery, Inc, pp. 55-61, 1st International Workshop on Adversarial Learning for Multimedia, AdvM 2021, co-located with ACM MM 2021, Virtual, Online, China, 20/10/21. https://doi.org/10.1145/3475724.3483609

An Investigation on Sparsity of CapsNets for Adversarial Robustness. / Zhao, Lei; Huang, Lei.
AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021. Association for Computing Machinery, Inc, 2021. p. 55-61 (AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021).

Research output: Chapter in Book/Report/Conference proceeding › Conference contribution › peer-review

TY - GEN

T1 - An Investigation on Sparsity of CapsNets for Adversarial Robustness

AU - Zhao, Lei

AU - Huang, Lei

PY - 2021/10/22

Y1 - 2021/10/22

N2 - The routing-by-agreement mechanism in capsule networks (CapsNets) is used to build visual hierarchical relationships with a characteristic of assigning parts to wholes. The connections between capsules of different layers become sparser with more iterations of routing. This paper proposes techniques in measuring, controlling, and visualizing the sparsity of CapsNets. One essential observation in this paper is that the sparser CapsNets are possibly more robust to the adversarial attacks. We believe this observation will provide insights into designing more robust models.

AB - The routing-by-agreement mechanism in capsule networks (CapsNets) is used to build visual hierarchical relationships with a characteristic of assigning parts to wholes. The connections between capsules of different layers become sparser with more iterations of routing. This paper proposes techniques in measuring, controlling, and visualizing the sparsity of CapsNets. One essential observation in this paper is that the sparser CapsNets are possibly more robust to the adversarial attacks. We believe this observation will provide insights into designing more robust models.

KW - adversarial robustness

KW - capsnet

KW - class activation maps

KW - sparsity

UR - https://www.scopus.com/pages/publications/85121109079

U2 - 10.1145/3475724.3483609

DO - 10.1145/3475724.3483609

M3 - 会议稿件

AN - SCOPUS:85121109079

T3 - AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021

SP - 55

EP - 61

BT - AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021

PB - Association for Computing Machinery, Inc

T2 - 1st International Workshop on Adversarial Learning for Multimedia, AdvM 2021, co-located with ACM MM 2021

Y2 - 20 October 2021 through 20 October 2021

ER -

Zhao L, Huang L. An Investigation on Sparsity of CapsNets for Adversarial Robustness. In AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021. Association for Computing Machinery, Inc. 2021. p. 55-61. (AdvM 2021 - Proceedings of the 1st International Workshop on Adversarial Learning for Multimedia, co-located with ACM MM 2021). doi: 10.1145/3475724.3483609

An Investigation on Sparsity of CapsNets for Adversarial Robustness

Abstract

Publication series

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Cite this