MoEvA2 Replication package

Important notice: The goal of this package is the replication of the experiments of our paper A Unified Framework for Adversarial Attack and Defense in Constrained Feature Space. This repository will not be updated. The latest version of our framework is available at https://github.com/serval-uni-lu/constrained-attacks.

Introduction

This package contains all the necessary data and scripts to replicate the experiments of our research paper "A Unified Framework for Adversarial Attack and Defense in Constrained Feature Space".

Create the environment

We recommend using miniconda and Python 3.8.8 on a Linux system.

conda create -n moeva2 python=3.8.8
conda activate moeva2
pip install -r requirements.txt

Download the additional data from https://figshare.com/s/84ae808ce6999fafd192 and place the content of the downloaded data folder in the data folder of this folder.

Alternatively, run the script './prepare.sh' to create the conda enviornment and place the data automatically.

File structure

• config: configurations files used to prepare and run the experiments
• data: the source data used to craft the attacks
• models: pre-trained models
• out: results of the experiments
• src: scripts and modules to run the attacks
  • attacks: module containing the implementations of the attacks
  • examples: usecase-specific scripts
  • botnet and lcld: scripts to prepare data and models for the experiments for CTU-13 and LCLD respectively
  • united: scripts to run a single attack

Usage

Run the following command to launch all the experiments at once.

conda activate moeva2
./run_all.sh

The results are found in ./out/attacks/[project]/rq[X]/metrics_[attack_name]_[hash].json

Feel free to comment out experiments in run_all.sh and/or reduce the number of original examples (n_initial_state parameter) in the config file ending with static.yaml to speed up the experiments.

The data used to train the models are also provided in the archive data-1 and data-2. Simply merge the folders if you intend to use them.

Use the following commands to retrain the models:

python -m src.experiments.lcld.01_train_robust -c ./config/rq0.lcld.yaml
python -m src.experiments.botnet.01_train_robust -c ./config/rq0.botnet.yaml

Citation

If you have used our framework for research purposes, you can cite our publication by:

BibTex:

@article{simonetto2021unified,
  title={A unified framework for adversarial attack and defense in constrained feature space},
  author={Simonetto, Thibault and Dyrmishi, Salijona and Ghamizi, Salah and Cordy, Maxime and Traon, Yves Le},
  journal={arXiv preprint arXiv:2112.01156},
  year={2021}
}