thuwzy / cluster-attack

Cluster Attack: Query-based Adversarial Attacks on Graphs with Graph-Dependent Priors (IJCAI 2022 Long Presentation)

Cluster Attack

The official implementation of paper Cluster Attack: Query-based Adversarial Attacks on Graphs with Graph-Dependent Priors (IJCAI 2022 Long Presentation).

Usage

To run Cluster Attack on discrete feature space with default hyper-parameters

python main.py

To run Cluster Attack on continuous feature space with default hyper-parameters

python main_concrete.py

Datasets

Two tiny citation networks (Cora and Citeseer) are included in our code.

Reddit dataset: matenure/FastGCN#8 (comment) (from FastGCN, a preprocessed version) or http://snap.stanford.edu/graphsage/ (from SNAP, original).

ogbn-arxiv dataset: https://snap.stanford.edu/ogb/data/nodeproppred/arxiv.zip

We provide checkpoints of GCN model on Reddit and ogbn-arxiv datasets.

Citation

If you find Cluster Attack helpful, please cite our paper.

@inproceedings{
    wang2022cluster,
    title={Cluster Attack: Query-based Adversarial Attacks on Graphs with Graph-Dependent Priors},
    author={Wang, Zhengyi and Hao, Zhongkai and Wang, Ziqiao and Su, Hang and Zhu, Jun},
    booktitle={International Joint Conference on Artificial Intelligence},
    year={2022},
    url={https://arxiv.org/abs/2109.13069}
}

Reference

This implementation is based on the following repos:

• TDGIA
• GraphSAGE