Codebase used to generated the results for the work Provable Robustness of (Graph) Neural Networks Against Data Poisoning and Backdoor Attacks. A preliminary version appeared at the AdvML-Frontiers @ NeurIPS 2024 workshop where it won the Best Paper Award and the full paper can be found on arXiv.
The codebase has been developed using Python 3.11 and requires the following packages and has been tested with the listed versions:
- pytorch 2.0.1
- pyg (pytorch geometric) 2.3.1
- torch-scatter 2.1.1.+pt20cu118
- torch-sparse 0.6.17+pt20cu118
- qpth 0.0.16
- gurobi 11.0.1
- jaxtyping 0.2.22
- typeguard 4.1.4
- networkx 3.1
- numpy 1.25.2
- seml 0.4.0
- scikit-learn 1.2.2
- cvxopt 1.3.2
- jupyterlab 4.0.6
- pandas 2.1.3
Experiments were run using the seml framework and a corresponding MongoDB installation (see seml-link). However, they can be run independently of seml and a MongoDB installation using the corresponding jupyter notebooks provided in the root directory and prefixed with test_.
We define the following experiments:
Jupyter Notebook: test_certify.ipynb
seml details:
- Experiment source file:
exp_certify.py - Experiment configurations:
config/certification/
Jupyter Notebook: test_certify_multiclass.ipynb
seml details:
- Experiment source file:
exp_certify_multiclass.py - Experiment configurations:
config/certification/
Jupyter Notebook: exp_certify_backdoor.ipynb
seml details:
- Experiment source file:
exp_certify_backdoor.py - Experiment configurations:
config/certification/
Jupyter Notebook: test_certify_backdoor_csbm.ipynb
seml details:
- Experiment source file:
exp_certify_backdoor_csbm.py - Experiment configurations:
config/certification/
Jupyter Notebook: test_attack.ipynb
seml details:
- Experiment source file:
exp_attack.py - Experiment configurations:
config/attack/
Jupyter Notebook: test_hyperpar.ipynb
seml details:
- Experiment source file:
exp_hyperparam.py - Experiment configurations:
config/hyperparams/
You can cite our full paper as follows:
@article{gosch2024provable,
title={Provable Robustness of (Graph) Neural Networks Against Data Poisoning and Backdoor Attacks},
author={Gosch, Lukas and Sabanayagam, Mahalakshmi and Ghoshdastidar, Debarghya and G{\"u}nnemann, Stephan},
journal={arXiv preprint arXiv:2407.10867},
year={2024}
}
To cite the AdvML-Frontiers @ NeurIPS24 workshop paper, please use
@article{gosch2024provable,
title={Provable Robustness of (Graph) Neural Networks Against Data Poisoning and Backdoor Attacks},
author={Gosch, Lukas and Sabanayagam, Mahalakshmi and Ghoshdastidar, Debarghya and G{\"u}nnemann, Stephan},
journal={The 3rd Workshop on New Frontiers in Adversarial Machine Learning, NeurIPS},
year={2024}
}
For questions and feedback, feel free to contact
Lukas Gosch, lukas (dot) gosch (at) tum (dot) de
Mahalakshmi Sabanayagam, m (dot) sabanayagam (at) tum (dot) de
This codebase contains code fragments from
We thank the authors for making their code publicly available.