A PyTorch implementation of "Graph Structure Learning for Robust Graph Neural Networks" (KDD 2020). [paper] [slides]

The code is based on our Pytorch adversarial repository, DeepRobust (https://github.com/DSE-MSU/DeepRobust)

Graph Neural Networks (GNNs) are powerful tools in representation learning for graphs. However, recent studies show that GNNs are vulnerable to carefully-crafted perturbations, called adversarial attacks. Adversarial attacks can easily fool GNNs in making predictions for downstream tasks. The vulnerability to adversarial attacks has raised increasing concerns for applying GNNs in safety-critical applications. Therefore, developing robust algorithms to defend adversarial attacks is of great significance. A natural idea to defend adversarial attacks is to clean the perturbed graph. It is evident that real-world graphs share some intrinsic properties. For example, many real-world graphs are low-rank and sparse, and the features of two adjacent nodes tend to be similar. In fact, we find that adversarial attacks are likely to violate these graph properties. Therefore, in this paper, we explore these properties to defend adversarial attacks on graphs. In particular, we propose a general framework Pro-GNN, which can jointly learn a structural graph and a robust graph neural network model from the perturbed graph guided by these properties. Extensive experiments on real-world graphs demonstrate that the proposed framework achieves significantly better performance compared with the state-of-the-art defense methods, even when the graph is heavily perturbed.

See that in https://github.com/DSE-MSU/DeepRobust/blob/master/requirements.txt

matplotlib==3.1.1
numpy==1.17.1
torch==1.2.0
scipy==1.3.1
torchvision==0.4.0
texttable==1.6.2
networkx==2.4
numba==0.48.0
Pillow==7.0.0
scikit_learn==0.22.1
skimage==0.0
tensorboardX==2.0

To run the code, first you need to clone DeepRobust

git clone https://github.com/DSE-MSU/DeepRobust.git
cd DeepRobust
python setup.py install

After installation, you can clone this repository

git clone https://github.com/ChandlerBang/Pro-GNN.git
cd Pro-GNN
python train.py --dataset polblogs --attack meta --ptb_rate 0.15 --epoch 1000

All the hyper-parameters settings are included in scripts folder. Note that same hyper-parameters are used under different perturbation for the same dataset.

To reproduce the performance reported in the paper, you can run the bash files in folder scripts.

sh scripts/meta/cora_meta.sh

IMPORTANT NOTICE For the performance of Pubmed dataset, if you don't add the code (line 59-62 in train.py), the performance of GCN should be around 85 since the data splits are different from what I used in the paper. See details in ChandlerBang#2

With the help of DeepRobust, you can run the following code to generate meta attack

python generate_attack.py --dataset cora --ptb_rate 0.05 --seed 15

For more information, you can take a look at the paper or the detailed code shown in DeepRobust.

If you find this repo to be useful, please cite our paper. Thank you.

@article{jin2020graph,
  title={Graph Structure Learning for Robust Graph Neural Networks},
  author={Jin, Wei and Ma, Yao and Liu, Xiaorui and Tang, Xianfeng and Wang, Suhang and Tang, Jiliang},
  journal={arXiv preprint arXiv:2005.10203},
  year={2020}
}