AdvGrads

🌐 About

AdvGrads is an all-in-one tool for comprehensive experimentation with adversarial attacks on image recognition. This repository provides an environment used for research purposes to validate the performance of attack and defense methods in adversarial attacks.

This repository is still under development. For more information, please contact with me (m.yuito3@gmail.com).

💻 Installation

Create environment

AdvGrads requires Python >= 3.9. The following is an example of building an environment using conda:

conda create --name advgrads -y python=3.9
conda activate advgrads
pip install --upgrade pip

Dependencies

Install other packages including PyTorch with CUDA (this repo has been tested with CUDA 11.8).

pip install torch==2.1.2+cu118 torchvision==0.16.2+cu118 --extra-index-url https://download.pytorch.org/whl/cu118
pip install -r requirements.txt

Installing AdvGrads

Install AdvGrads in editable mode from source code:

python -m pip install -e .

🚀 Usage

You can execute the attack immediately using the config files provided in this repository.

python advgrads_cli/attack.py --load_config configs/mnist_untargeted.yaml

⚙ Description format of config files

The attack configs are managed by a YAML file. The main fields and variables are described below.

• data: (str, required) Specify a dataset for which adversarial examples are to be generated.
• model: (str, required) Select a model to be attacked. See here for currently supported models.
• attacks: (list, required) This field allows you to specify attack methods that you wish to execute in a list format. You can set values including hyperparameters defined for each method. The parameters that can be specified for all methods are as follows:
  • method: (str) Attack method. See here for currently supported attack methods.
  • norm: (str) Norm for adversarial perturbations.
  • eps: (float) Maximum norm constraint.
  • max_iters: (int) Maximum number of iterations used in iterative methods.
  • targeted: (bool) Whether or not to perform targeted attacks which aim to misclassify an adversarial example into a particular class.
• thirdparty_defense: (str, optional) Thirdparty defensive method. See here for currently supported defensive methods.

🔍 Features

AdvGrads is developed on PyTorch.

💥 Attacks

Currently supported attack methods are as follows:

Method	Type
DeepFool	White-box
DI-MI-FGSM	White-box
FGSM	White-box
I-FGSM (BIM)	White-box
MI-FGSM (MIM)	White-box
NI-FGSM	White-box
PGD	White-box
PI-FGSM	White-box
SI-NI-FGSM	White-box
SignHunter	Black-box
SimBA	Black-box
Square attack	Black-box

💠 Defenses

Currently supported defense methods including adversarially trained models are as follows:

Method	Type
Bit-Red	Input transform
JPEG	Input transform
Randomization	Input transform
TRADES	Adv. training

🧩 Others

And also, some naturally trained models are supported.

Source	Datasets
pytorch-playground	MNIST, CIFAR-10