Please cite:
@article{py3fst,
title={Cortical Features for Defense Against Adversarial Audio Attacks},
author={Ilya Kavalerov and Frank Zheng and Wojciech Czaja and Rama Chellappa},
journal={arXiv preprint},
year={2021}
}
See scripts
directory.
scripts/wake/data.sh
You will see data refered to in the scripts and ipython notebooks. v7.18
is the version of the data used in reported results (its the gpu version of v7.17
), ignore earlier versions.
scripts/wake_final/vulcan_baseline_train.sh
scripts/wake_final/vulcan_cortical_train.sh
scripts/wake_final/vulcan_baseline_eval.sh
scripts/wake_final/attack_univ_all.sh
To do this attack you will need to export a model of a different length first using:
scripts/wake/vulcan_export.124.sh
Then attack with:
scripts/wake_final/attack_univ_music.sh
scripts/wake_final/attack_univ_music_eval.sh
conda env create -f venvtf1p15nb.yml
Running on: Python 3.6.9, tensorflow 1.15, cuda/10.0.130, cudnn/v7.6.5.
For creating datasets, works with: ffmpeg/4.2.1, rubberband 1.8.2. Also using vamp-plugin-sdk 2.9 2019-11-13, libsndfile Version 1.0.28, libsndfile Version 0.1.9.
MIT