This repository contains the [PyTorch] implementation of the paper "Towards Frequency-Based Explanation for Robust CNN" arxiv. It primarly deals the extent to which image features are robust in the frequency domain.
- It is recommended to setup a fresh virtual environment first.
python -m venv env
source activate env/bin/activate
- Install the torchattacks package
pip install torchattacks
- Run the
main.py
file.
The original paper implemented the attacks using a VGG 19 model. However, due to memory constraints I did it using ResNet 18. Here I have provided a fine-tuned (on CIFAR 10) version of ResNet 18 which is pre-trained on ImageNet. The checkpoint can be downloaded using this link.
CW | FGSM | PGD |
---|---|---|
This is not an official implementation of the paper. I am not associated with the authors of the paper or Lab in any manner whatsoever and I don't claim credit for any of the algorithms proposed in the paper.