BackdoorVault is a toolbox that collects representative backdoor attacks. It is based on PyTorch and still under development.
Some part of the implementation is drawn from the original backdoor papers. Thank authors for their open-sourced implemetations.
The code is implemented and tested on PyTorch. It runs on Python 3.6.9.
The main functions are located in main.py file. For a test drive, please use the following command:
python3 main.py --phase poisonThis will generate a backdoored model by a polygon trigger with target label 0.
The following attacks have been included in this toolbox.
BadNets, Blend, DFST, Dynamic, Filter, Input-aware, Invisible, Refool, SIG, WaNet
If you find the toolbox useful for your research, please cite the following:
@article{taog2023backdoorvault,
title={{BackdoorVault}: A Toolbox for Backdoor Attacks},
author={Tao, Guanhong and Cheng, Siyuan},
year={2023}
}