Improving Robustness of DNNs against Common Corruptions via Gaussian Adversarial Training

Chenyu Yi; Haoliang Li; Renjie Wan; Alex C. Kot

doi:10.1109/VCIP49819.2020.9301856

Improving Robustness of DNNs against Common Corruptions via Gaussian Adversarial Training

Chenyu Yi, Haoliang Li, Renjie Wan, Alex C. Kot

Department of Computer Science

Research output: Chapter in book/report/conference proceeding › Conference proceeding › peer-review

2 Citations (Scopus)

Abstract

Deep neural networks have demonstrated tremendous success in image classification, but their performance sharply degrades when evaluated on slightly different test data (e.g., data with corruptions). To address these issues, we propose a minimax approach to improve common corruption robustness of deep neural networks via Gaussian Adversarial Training. To be specific, we propose to train neural networks with adversarial examples where the perturbations are Gaussian-distributed. Our experiments show that our proposed GAT can improve neural networks' robustness to noise corruptions more than other baseline methods. It also outperforms the state-of-the-art method in improving the overall robustness to common corruptions.

Original language	English
Title of host publication	2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020
Publisher	IEEE
Pages	17-20
Number of pages	4
ISBN (Electronic)	9781728180670, 9781728180687
ISBN (Print)	9781728180694
DOIs	https://doi.org/10.1109/VCIP49819.2020.9301856
Publication status	Published - 1 Dec 2020
Event	2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020 - Virtual, Macau, China Duration: 1 Dec 2020 → 4 Dec 2020 https://ieeexplore.ieee.org/xpl/conhome/9301747/proceeding

Publication series

Name	IEEE Visual Communications and Image Processing (VCIP)
ISSN (Print)	1018-8770
ISSN (Electronic)	2642-9357

Conference

Conference	2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020
Country/Territory	China
City	Macau
Period	1/12/20 → 4/12/20
Internet address	https://ieeexplore.ieee.org/xpl/conhome/9301747/proceeding

Scopus Subject Areas

Computer Networks and Communications
Signal Processing
Media Technology

User-Defined Keywords

Adversarial Training
Data Augmentation
Deep Learning
Robustness to Common Corruptions

Access to Document

10.1109/VCIP49819.2020.9301856

Cite this

@inproceedings{409b0c3da2b84451a80a9cde42cdf313,

title = "Improving Robustness of DNNs against Common Corruptions via Gaussian Adversarial Training",

abstract = "Deep neural networks have demonstrated tremendous success in image classification, but their performance sharply degrades when evaluated on slightly different test data (e.g., data with corruptions). To address these issues, we propose a minimax approach to improve common corruption robustness of deep neural networks via Gaussian Adversarial Training. To be specific, we propose to train neural networks with adversarial examples where the perturbations are Gaussian-distributed. Our experiments show that our proposed GAT can improve neural networks' robustness to noise corruptions more than other baseline methods. It also outperforms the state-of-the-art method in improving the overall robustness to common corruptions.",

keywords = "Adversarial Training, Data Augmentation, Deep Learning, Robustness to Common Corruptions",

author = "Chenyu Yi and Haoliang Li and Renjie Wan and Kot, {Alex C.}",

note = "Funding Information: The research work was done at the Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University, Singapore. The work is supported in part by the NTU-PKU Joint Research Institute, a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation. Renjie Wan and Haoliang Li are supported by Wallenberg-NTU Presidential Postdoctoral Fellowship. Publisher Copyright: {\textcopyright} 2020 IEEE.; 2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020 ; Conference date: 01-12-2020 Through 04-12-2020",

year = "2020",

month = dec,

day = "1",

doi = "10.1109/VCIP49819.2020.9301856",

language = "English",

isbn = "9781728180694",

series = "IEEE Visual Communications and Image Processing (VCIP)",

publisher = "IEEE",

pages = "17--20",

booktitle = "2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020",

address = "United States",

url = "https://ieeexplore.ieee.org/xpl/conhome/9301747/proceeding",

}

Yi, C, Li, H, Wan, R & Kot, AC 2020, Improving Robustness of DNNs against Common Corruptions via Gaussian Adversarial Training. in 2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020. IEEE Visual Communications and Image Processing (VCIP), IEEE, pp. 17-20, 2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020, Macau, China, 1/12/20. https://doi.org/10.1109/VCIP49819.2020.9301856

Improving Robustness of DNNs against Common Corruptions via Gaussian Adversarial Training. / Yi, Chenyu; Li, Haoliang; Wan, Renjie et al.
2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020. IEEE, 2020. p. 17-20 (IEEE Visual Communications and Image Processing (VCIP)).

Research output: Chapter in book/report/conference proceeding › Conference proceeding › peer-review

TY - GEN

T1 - Improving Robustness of DNNs against Common Corruptions via Gaussian Adversarial Training

AU - Yi, Chenyu

AU - Li, Haoliang

AU - Wan, Renjie

AU - Kot, Alex C.

N1 - Funding Information: The research work was done at the Rapid-Rich Object Search (ROSE) Lab, Nanyang Technological University, Singapore. The work is supported in part by the NTU-PKU Joint Research Institute, a collaboration between the Nanyang Technological University and Peking University that is sponsored by a donation from the Ng Teng Fong Charitable Foundation. Renjie Wan and Haoliang Li are supported by Wallenberg-NTU Presidential Postdoctoral Fellowship. Publisher Copyright: © 2020 IEEE.

PY - 2020/12/1

Y1 - 2020/12/1

N2 - Deep neural networks have demonstrated tremendous success in image classification, but their performance sharply degrades when evaluated on slightly different test data (e.g., data with corruptions). To address these issues, we propose a minimax approach to improve common corruption robustness of deep neural networks via Gaussian Adversarial Training. To be specific, we propose to train neural networks with adversarial examples where the perturbations are Gaussian-distributed. Our experiments show that our proposed GAT can improve neural networks' robustness to noise corruptions more than other baseline methods. It also outperforms the state-of-the-art method in improving the overall robustness to common corruptions.

AB - Deep neural networks have demonstrated tremendous success in image classification, but their performance sharply degrades when evaluated on slightly different test data (e.g., data with corruptions). To address these issues, we propose a minimax approach to improve common corruption robustness of deep neural networks via Gaussian Adversarial Training. To be specific, we propose to train neural networks with adversarial examples where the perturbations are Gaussian-distributed. Our experiments show that our proposed GAT can improve neural networks' robustness to noise corruptions more than other baseline methods. It also outperforms the state-of-the-art method in improving the overall robustness to common corruptions.

KW - Adversarial Training

KW - Data Augmentation

KW - Deep Learning

KW - Robustness to Common Corruptions

UR - http://www.scopus.com/inward/record.url?scp=85099485333&partnerID=8YFLogxK

U2 - 10.1109/VCIP49819.2020.9301856

DO - 10.1109/VCIP49819.2020.9301856

M3 - Conference proceeding

AN - SCOPUS:85099485333

SN - 9781728180694

T3 - IEEE Visual Communications and Image Processing (VCIP)

SP - 17

EP - 20

BT - 2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020

PB - IEEE

T2 - 2020 IEEE International Conference on Visual Communications and Image Processing, VCIP 2020

Y2 - 1 December 2020 through 4 December 2020

ER -

Improving Robustness of DNNs against Common Corruptions via Gaussian Adversarial Training

Abstract

Publication series

Conference

Scopus Subject Areas

User-Defined Keywords

Access to Document

Other files and links

Fingerprint

Cite this