CIFS: Improving Adversarial Robustness of CNNs via Channel-wise Importance-based Feature Selection

Hanshu Yan, Jingfeng Zhang, Gang Niu, Jiashi Feng, Vincent Tan, Masashi Sugiyama
Proceedings of the 38th International Conference on Machine Learning, PMLR 139:11693-11703, 2021.

Abstract

We investigate the adversarial robustness of CNNs from the perspective of channel-wise activations. By comparing normally trained and adversarially trained models, we observe that adversarial training (AT) robustifies CNNs by aligning the channel-wise activations of adversarial data with those of their natural counterparts. However, the channels that are \textit{negatively-relevant} (NR) to predictions are still over-activated when processing adversarial data. Besides, we also observe that AT does not result in similar robustness for all classes. For the robust classes, channels with larger activation magnitudes are usually more \textit{positively-relevant} (PR) to predictions, but this alignment does not hold for the non-robust classes. Given these observations, we hypothesize that suppressing NR channels and aligning PR ones with their relevances further enhances the robustness of CNNs under AT. To examine this hypothesis, we introduce a novel mechanism, \textit{i.e.}, \underline{C}hannel-wise \underline{I}mportance-based \underline{F}eature \underline{S}election (CIFS). The CIFS manipulates channels’ activations of certain layers by generating non-negative multipliers to these channels based on their relevances to predictions. Extensive experiments on benchmark datasets including CIFAR10 and SVHN clearly verify the hypothesis and CIFS’s effectiveness of robustifying CNNs.

Cite this Paper

BibTeX
@InProceedings{pmlr-v139-yan21e, title = {CIFS: Improving Adversarial Robustness of CNNs via Channel-wise Importance-based Feature Selection}, author = {Yan, Hanshu and Zhang, Jingfeng and Niu, Gang and Feng, Jiashi and Tan, Vincent and Sugiyama, Masashi}, booktitle = {Proceedings of the 38th International Conference on Machine Learning}, pages = {11693--11703}, year = {2021}, editor = {Meila, Marina and Zhang, Tong}, volume = {139}, series = {Proceedings of Machine Learning Research}, month = {18--24 Jul}, publisher = {PMLR}, pdf = {http://proceedings.mlr.press/v139/yan21e/yan21e.pdf}, url = {https://proceedings.mlr.press/v139/yan21e.html}, abstract = {We investigate the adversarial robustness of CNNs from the perspective of channel-wise activations. By comparing normally trained and adversarially trained models, we observe that adversarial training (AT) robustifies CNNs by aligning the channel-wise activations of adversarial data with those of their natural counterparts. However, the channels that are \textit{negatively-relevant} (NR) to predictions are still over-activated when processing adversarial data. Besides, we also observe that AT does not result in similar robustness for all classes. For the robust classes, channels with larger activation magnitudes are usually more \textit{positively-relevant} (PR) to predictions, but this alignment does not hold for the non-robust classes. Given these observations, we hypothesize that suppressing NR channels and aligning PR ones with their relevances further enhances the robustness of CNNs under AT. To examine this hypothesis, we introduce a novel mechanism, \textit{i.e.}, \underline{C}hannel-wise \underline{I}mportance-based \underline{F}eature \underline{S}election (CIFS). The CIFS manipulates channels’ activations of certain layers by generating non-negative multipliers to these channels based on their relevances to predictions. Extensive experiments on benchmark datasets including CIFAR10 and SVHN clearly verify the hypothesis and CIFS’s effectiveness of robustifying CNNs.} }
Endnote
%0 Conference Paper %T CIFS: Improving Adversarial Robustness of CNNs via Channel-wise Importance-based Feature Selection %A Hanshu Yan %A Jingfeng Zhang %A Gang Niu %A Jiashi Feng %A Vincent Tan %A Masashi Sugiyama %B Proceedings of the 38th International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2021 %E Marina Meila %E Tong Zhang %F pmlr-v139-yan21e %I PMLR %P 11693--11703 %U https://proceedings.mlr.press/v139/yan21e.html %V 139 %X We investigate the adversarial robustness of CNNs from the perspective of channel-wise activations. By comparing normally trained and adversarially trained models, we observe that adversarial training (AT) robustifies CNNs by aligning the channel-wise activations of adversarial data with those of their natural counterparts. However, the channels that are \textit{negatively-relevant} (NR) to predictions are still over-activated when processing adversarial data. Besides, we also observe that AT does not result in similar robustness for all classes. For the robust classes, channels with larger activation magnitudes are usually more \textit{positively-relevant} (PR) to predictions, but this alignment does not hold for the non-robust classes. Given these observations, we hypothesize that suppressing NR channels and aligning PR ones with their relevances further enhances the robustness of CNNs under AT. To examine this hypothesis, we introduce a novel mechanism, \textit{i.e.}, \underline{C}hannel-wise \underline{I}mportance-based \underline{F}eature \underline{S}election (CIFS). The CIFS manipulates channels’ activations of certain layers by generating non-negative multipliers to these channels based on their relevances to predictions. Extensive experiments on benchmark datasets including CIFAR10 and SVHN clearly verify the hypothesis and CIFS’s effectiveness of robustifying CNNs.
APA
Yan, H., Zhang, J., Niu, G., Feng, J., Tan, V. & Sugiyama, M.. (2021). CIFS: Improving Adversarial Robustness of CNNs via Channel-wise Importance-based Feature Selection. Proceedings of the 38th International Conference on Machine Learning, in Proceedings of Machine Learning Research 139:11693-11703 Available from https://proceedings.mlr.press/v139/yan21e.html.

Related Material