Graph Structure Reshaping Against Adversarial Attacks on Graph Neural Networks

Code implementation of the paper: Graph Structure Reshaping Against Adversarial Attacks on Graph Neural Networks, which has been submitted to TKDE for review.

Requirements

• python 3.6.7
• numpy 1.15.4
• scipy 1.1.0
• scikit-learn 0.20.2
• matplotlib 3.0.2
• torch 1.0.0
• tqdm 4.31.1

Hardware Configurations

All experiments are conducted on a server with the following configurations:

• Operating System: CentOS Linux release 7.4.1708
• CPU: Intel(R) Xeon(R) CPU E5-2620 v4 @ 2.20GHz
• GPU: GeForce GTX TITAN X

Run the code

To try our code, you can use the IPython notebook demo.ipynb.

Datasets

In the folder ./data, we provide the following datasets:

Clean Graph

cora.npz, citeseer.npz and coraml.npz

Poisoned Graph Generated by Meta-Self

[dataset]_[attack_ratio]edges_Meta-Self.npy

Poisoned Graph Generated by Meta-Train

[dataset]_[attack_ratio]edges_Meta-Train.npy

Poisoned Graph Generated by Min-Max Attack

[dataset]_[attack_ratio]_minmax.npy

Baselines and Adversarial Attack Methods

We use the following publicly available implementation of baseline methods and adversarial attack methods:

Attack Method

• Meta-Train && Meta-Self: https://github.com/danielzuegner/gnn-meta-attack
• Min-Max Attack: https://github.com/KaidiXu/GCN_ADV_Train
• NETTACK: https://github.com/danielzuegner/nettack

Defense Method

• RGCN: https://github.com/thumanlab/nrlweb/blob/master/static/assets/download/RGCN.zip
• VPN: https://www.dropbox.com/sh/p36pzx1ock2iamo/AABEr7FtM5nqwC4i9nICLIsta?dl=0
• LinkPred && SubGLinkPred: https://openreview.net/pdf?id=HJxdAoCcYX
• PreProcess: https://arxiv.org/pdf/1903.01610.pdf

Visualization Example Output