SGA

Official implementation of Simlified Gradient Attack (SGA) in TKDE 2021: Adversarial attack on large scale graph.

Implementations

• The PyTorch implementation can be found at DeepRobust or GraphWar.

Update

• 2021/3/20: add tensorflow 2.x support

Requirements

• TensorFlow 1.15 (it should work in 1.14, 1.13... and 2.x)
• Numpy
• Scipy
• tqdm

Datasets

we consider the largest connected component (LCC) of the graph for each dataset, e.g., data/cora_lcc.

Cora, Citeseer and Pubmed could be found at https://github.com/danielzuegner/gnn-meta-attack,

reddit could be found at http://snap.stanford.edu/graphsage/

• cora
• citeseer
• pubmed
• reddit

Usage

How to attack a graph neural networks?

• Train the surrogate model SGC
• Generate perturbation (edge flips)
• Attack target models with adversarial examples (SGC, GCN ...)

See src/demo.ipynb for a simple example to run our codes.

Reproduce results in paper

• Run src/generate_edges.ipynb to generate adversarial edges and save them, the adversarial edges can be generated by SGA, GradArgmax, Nettack.
• Run src/attack.ipynb to test the attack performance on SGC and GCN.

Cite

If you make use of this code or the SGA algorithm in your work, please cite the following paper:

@article{li2021adversarial,
  title={Adversarial attack on large scale graph},
  author={Li, Jintang and Xie, Tao and Liang, Chen and Xie, Fenfang and He, Xiangnan and Zheng, Zibin},
  journal={IEEE Transactions on Knowledge and Data Engineering},
  year={2021},
  publisher={IEEE}
}