Skip to content

Latest commit

 

History

History

defense

README.md

Requirements:

  • easydict for AT
  • statsmodels for RS
  • opencv-python for NRP

Quick Start:

  1. Please change the name of the "ATTACK_METHOD" in '.sh' files to evaluate your method!
  2. Directly run the '.sh' files according to the following instructions!

Notes: We've already set up the relatively path to run, so you can run these commands directly without making many changes!

AT: https://github.com/locuslab/fast_adversarial/tree/master/ImageNet

sh at_defense.sh

We use default 4eps evaluation.

HGD: https://github.com/lfz/Guided-Denoise/tree/master/nips_deploy

sh hgd_defense.sh

RS: https://github.com/locuslab/smoothing

sh rs_defense.sh

python defense/rs/predict.py /path/to/adv_data /path/to/noise_0.50/checkpoint.pth.tar  0.50 prediction_outupt --alpha 0.001 --N 1000 --skip 100 --batch 1

Time: ~ 1 hour for 1000 samples on a single 4090 GPU

NRP: https://github.com/Muzammal-Naseer/NRP

sh nrp_defense.sh

python defense/nrp/purify.py --dir=/path/to/adv_data --output=/path/to/save/purified_data --purifier NRP --model_pth /path/to/NRP.pth  --dynamic

Then, evaluate the purified_data, we report the ASR on ResNet101 target model.

DiffPure: https://github.com/NVlabs/DiffPure

sh diffpure_defense.sh

Then, evaluate the purified_data, we report the ASR on ResNet101 target model.