README

This is for releasing the source code of the paper "Februus: Input Purification Defense Against Trojan Attacks on Deep Neural Network Systems"

Archived Version: Februus

The project is published as part of the following paper and if you re-use our work, please cite the following paper:

@inproceedings{doan2020februus,
title={Februus: Input Purification Defense Against Trojan Attacks on Deep Neural Network Systems},
author={Bao Gia Doan and Ehsan Abbasnejad and Damith C. Ranasinghe},
year = {2020},
booktitle = {Proceedings of the 36th Annual Computer Security Applications Conference (ACSAC)},
location = {Austin, TX, USA},
series = {ACSAC 2020}
}

The source code is written mostly on Python 3 and Pytorch, so please help to download and install Python3 and Pytorch beforehand.

There are some large files in the repo, so it is recommended to install git LFS:

sudo apt-get install git-lfs
git lfs install

Install cmake (optionally):

sudo apt-get install cmake

Requirements

To install the requirements for this repo, run the following command:

git clone https://github.com/AdelaideAuto-IDLab/Februus.git
cd Februus
pip3 install -r requirements.txt
pip3 install torch==1.6.0+cu101 torchvision==0.7.0+cu101 -f https://download.pytorch.org/whl/torch_stable.html

Large Files and Datasets

For the convenience of the users, we put the GTSRB and VGGFace2-val set that we used in datasets folder. There are some large files but with the limited storage of Github LFS, we put large files on a separate link, please help to follow the below instructions:

• Download the largefile file (https://universityofadelaide.box.com/s/u7di8pcirjc1flvnzwcu16l2ddujorb8)

• Download the BTSR file for Belgium Traffic Sign Recognition (https://universityofadelaide.box.com/s/wo567ru7tuxrcfjz2ypbkarqat87eabw)

• Put the largefile.tar.gz and BTSR.tar.gz to the root of Februus repo folder (i.e. Februus/largefile.tar.gz)

• Untar the files

tar -xzvf largefile.tar.gz
tar -xzvf BTSR.tar.gz

• Untar the dataset within the datasets folder (please don't change the name of the folder, or you'll need to adjust the location of the dataset later in the source code).

cd datasets
tar -xzvf gtsrb-dataset.tar.gz
tar -xzvf vggface2-val.tar.gz

Run the Februus

There are 4 sub-repo, each for a different visual classification task:

• face: Face Recognition Task (VGGFace2)
• scene: Scene Classification Task (CIFAR-10)
• traffic_sign: Traffic Sign Recognition Task (GTSRB)
• BTSR: Belgium Traffic Sign Recognition (BTSR)

There are two ways to run the method:

• The first way is to run step-by-step with the Jupyter Notebook file Februus.ipynb inside of each repo.

• The second way is to run the Februus.py file within each sub-repo. This is to run the Februus on the whole test set for that task.:

# For example, to run Februus on Scene Classification task
cd scene
python3 Februus.py

TODO

• add the training code