Skip to content

Official Respository of the paper A Hierarchical Assessment of Adversarial Severity (Best Paper Award at ICCV2021 Workshop AROW)

License

Notifications You must be signed in to change notification settings

BCV-Uniandes/AdvSeverity

Repository files navigation

Official Repository: A Hierarchical Assessment of Adversarial Severity

This is the official repository for the paper A Hierarchical Assessment of Adversarial Severity. This paper recieved the Best Paper Awards at the ICCV2021 workshop on Adversarial Robustness in the Real World)

Hierarchical Adversarial Attacks

Setting up the environment

First, install the environment via anaconda by following:

conda env create -f environment.yml
conda activate AdvSeverity

Then, install AutoAttack and tqdm by running:

pip install git+https://github.com/fra31/auto-attack
pip install tqdm

Data preparation

  • Download train+val sets of iNaturalist'19
  • Unzip the zip file dataset_splits/splits_inat19.py
  • Create the dataset by running the generate_splits.py script. Change the DATA_PATH AND OUTPUT_DIR variables to fit your specifications.
  • Resize all the images into the 224x224 format by using the resize_images.sh script: bash ./resize_images.sh new_splits_backup new_splits_downsampled 224x224!
  • Rename data_paths.yml and edit it to reflect the paths on your system.

Model Downloads

You can download our models directly from this link

Training and Evaluation

Training

To run the training rutine, run the main.py script as follows:

python main.py \
    --arch resnet18 \
    --dropout 0.5 \
    --output PATH/TO/OUTPUT \
    --num-training-steps 200000 \
    --gpu GPU \
    --val-freq 1 \
    --attack-eps EPSILON \
    --attack-iter-training ITERATIONS \
    --attack-step STEP \
    --attack free \
    --curriculum-training

If you want to train the method without the proposed curriculum, don't use the flag --curriculum-training. Further, if you want to train the model with TRADES change the flag --attack free with --attack trades.

To restore the training from a checkpoint, just use the same command the fristly used.

Evaluation

To perform the evaluation with the proposed attacks, pgd or NHAA, run the main.py script as follows:

python main.py \
    --arch resnet18 \
    --output PATH/TO/OUTPUT \
    --gpu GPU \
    --val-freq 1 \
    --attack-eps EPSILON \
    --attack-iter-training ITERATIONS \
    --attack-step 1 \
    --evaluate ATTACK \
    --attack-iter-evaluation ATTACKITERATIONS

If --evaluate uses as input hPGD, use the --hPGD flag to select between LHA, GHA or NHA and --hPGD-level to select the target height. To set the total number of attack iterations, use the flag --attack-iter-training.

Citation

If you found our paper or code useful, please cite our work:

@InProceedings{Jeanneret_2021_ICCV,
    author    = {Jeanneret, Guillaume and Perez, Juan C. and Arbelaez, Pablo},
    title     = {A Hierarchical Assessment of Adversarial Severity},
    booktitle = {Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops},
    month     = {October},
    year      = {2021},
    pages     = {61-70}
}

This code was based on Bertinetto's Making Better Mistakes official repository

License

This work is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License. Commercial licenses available upon request.

About

Official Respository of the paper A Hierarchical Assessment of Adversarial Severity (Best Paper Award at ICCV2021 Workshop AROW)

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published