README.md

Aliasing is a Driver of Adversarial Attacks

This is the official implementation of Aliasing is a Driver of Adversarial Attacks.

In this work, we investigate the hypothesis that the existence of adversarial perturbations is due in part to aliasing in neural networks. We made heavy use of the fast up-sampling, down-sampling, and anti-aliased ReLu implementations of Stylegan3 [Karras et al].

[Webpage] [Paper] [arXiv]

Requirements

To use our code run the following:

git clone https://github.com/adriarm/aliasing_is_a_driver
cd aliasing_is_a_driver
pip install -r requirements.txt

Experiments

To re-run our experiments, you will need a distributed cluster running Slurm. The current slurm configuration requests 8 nodes with 1 GPU each (we used Nvidia Volta V100).

# Download Tiny-Imagenet
mkdir datasets
wget -P datasets/ http://cs231n.stanford.edu/tiny-imagenet-200.zip
unzip datasets/tiny-imagenet-200.zip

# Logs
mkdir slurm_logs # folder to deposit the logs

# Submit scripts to Slurm
sbatch ./sh/train/cifar10_vgg11.sh
sbatch ./sh/train/cifar10_resnet50.sh
sbatch ./sh/train/tinyimagenet_vgg11.sh
sbatch ./sh/train/tinyimagenet_resnet50.sh
sbatch ./sh/train/typewise.sh

If you want to use a different node/gpu configuration, you will need to change the resource request slurm commands in the .sh files, as well as the batch size used, accordingly. The default per-gpu batch size for 8 GPUs is found in the corresponding config file in conf/train/trainer.

# Slurm resources options
#SBATCH --nodes=8
#SBATCH --ntasks-per-node=1
#SBATCH --gres=gpu:volta:1
#SBATCH --cpus-per-task=8
.
.
.
srun python3 launch_train.py ... trainer.train_loader.batch_size=X trainer.train_loader.valid_size=X

Citation

  @misc{rodriguezmunoz2022driver,
    title = {Aliasing is a Driver of Adversarial Attacks},
    author = {Rodríguez-Muñoz, Adrián and Torralba, Antonio},
    year={2022},
    eprint={2106.05963},
    archivePrefix={arXiv},
    primaryClass={cs.CV},
    url = {https://arxiv.org/abs/2212.11760},
  }