CIFS

This repository provides codes for CIFS (ICML 2021).

CIFS: Improving Adversarial Robustness of CNNs via Channel-wise Importance-based Feature Selection
Hanshu Yan, Jingfeng Zhang, Gang Niu, Jiashi Feng, Vincent Y. F. Tan, Masashi Sugiyama

Requirements

Python = 3.6
Pytorch = 1.6
CUDA = 10.1
Advertorch = 0.2.3

Channel-wise Importance-based Feature Selection

<p align="center"> <img src="markdown/picture.png" width="500"\> </p> <p align="center"> CIFS: 1) Probe Network A^l first makes a raw prediction p^l for z^l. 2) Channels’ relevances g^l are assessed (Relev. Ass.) based on the gradients of the top-k prediction results y^{l,k}. 3) The IMGF generates an importance mask m^l from g^l for channel adjustment. </p>