| 講演抄録／キーワード |
| 講演名 |
2022-02-21 16:15
A Note on Realizing Adversarial Defense Based on Regularization of Multi-stage Squeeze-and-Excitation Features
○Jiahuan Zhang・Keisuke Maeda・Takahiro Ogawa・Miki Haseyama（Hokkaido Univ.） |
| 抄録 |
（和） |
Regularizing deep features is a common adversarial defense method. However, the existing methods do not further explore the effects of regularizing the features with strongly expressive ability on the adversarial robustness. Squeeze-and-Excitation (SE) is a kind of widely used attention modules, which can fully consider the dependencies among different channels so that the output features of SE (SE features) can learn the channel-level importance. On the other hand, the SE features have strong representativeness. Therefore, we regularize a new architecture that combines ResNet and SE in this paper. Extensive white-box attack experiments have fully demonstrated the robustness and effectiveness of the proposed method. Moreover, we also illustrate that regularizing the features with the strongly expressive ability can further improve the robustness. |
| （英） |
Regularizing deep features is a common adversarial defense method. However, the existing methods do not further explore the effects of regularizing the features with strongly expressive ability on the adversarial robustness. Squeeze-and-Excitation (SE) is a kind of widely used attention modules, which can fully consider the dependencies among different channels so that the output features of SE (SE features) can learn the channel-level importance. On the other hand, the SE features have strong representativeness. Therefore, we regularize a new architecture that combines ResNet and SE in this paper. Extensive white-box attack experiments have fully demonstrated the robustness and effectiveness of the proposed method. Moreover, we also illustrate that regularizing the features with the strongly expressive ability can further improve the robustness. |
| キーワード |
（和） |
Adversarial attack / adversarial defense / Squeeze-and-Excitation / ResNet / regularization / / / |
| （英） |
Adversarial attack / adversarial defense / Squeeze-and-Excitation / ResNet / regularization / / / |
| 文献情報 |
映情学技報, vol. 46, no. 6, ME2022-42, pp. 87-90, 2022年2月. |
| 資料番号 |
ME2022-42 |
| 発行日 |
2022-02-14 (MMS, ME, AIT) |
| ISSN |
Print edition: ISSN 1342-6893 Online edition: ISSN 2424-1970 |
| PDFダウンロード |
|
|