SPAC

This is the code for the paper "Graph Structural Attack by Perturbing Spectral Distance" accepted by KDD 2022.

Requirement

Code is tested in Python 3.10.10. Some major requirements are listed below:

pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117
pip install torch_geometric
pip install pyg_lib torch_scatter torch_sparse torch_cluster torch_spline_conv -f https://data.pyg.org/whl/torch-1.13.0+cu117.html
pip install dgl
pip install networkx
pip install numba

Run the code

To execute the attack, you will first train and save a clean model; then attack a loaded model in both training time (poisoning attack) and test time (evasion attack). The following example commands on Cora data are run under the root folder.

Train a Clean Model

python run_train_nat.py --dataset cora --gnn_epochs 200 --lr 0.01

By default, the clean model will be saved at ./log/nat_model_saved/cora_GCN.pt

Attack the Trained Model

When pairing SPAC with other attacks, the argument --spac_weight controls the strength of SPAC term, and --loss_weight controls the other attack's task loss term.

• For poisoning attack: --attacker can choose from [minmax, Meta-Self, Meta-Train, random]
• For evasion attack: --attacker can choose from [PGD, random]

Poisoning Attack

Run SPAC (spectral attack) alone:

python run_attack_poison.py --gnn_path ./log/nat_model_saved/cora_GCN.pt --spac_weight 1.0 --loss_weight 0.0 

Run SPAC-Min (Max-Min paired with SPAC):

python run_attack_poison.py --gnn_path ./log/nat_model_saved/cora_GCN.pt --spac_weight 1.0 --loss_weight 1.0 

Evasion Attack

Run SPAC (spectral attack) alone:

python run_attack_evasion.py --gnn_path ./log/nat_model_saved/cora_GCN.pt --spac_weight 1.0 --loss_weight 0.0 

Run SPAC-CE (PGD-CE paired with SPAC):

python run_attack_evasion.py --gnn_path ./log/nat_model_saved/cora_GCN.pt --spac_weight 1.0 --loss_weight 1.0

Cite

Please cite our paper if you find this repo useful for your research or development.

@inproceedings{lin2022graph,
  title={Graph structural attack by perturbing spectral distance},
  author={Lin, Lu and Blaser, Ethan and Wang, Hongning},
  booktitle={Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={989--998},
  year={2022}
}