GUARD: Graph Universal Adversarial Defense

PyTorch implementation of the CIKM 2023 paper: "GUARD: Graph Universal Adversarial Defense" [arXiv].

Fig. 1. An illustrative example of graph universal defense. The universal patch p can be applied to an arbitrary node (here v1) to protect it from adversarial targeted attacks by removing adversarial edges (if exist).

Requirements

• torch==1.9
• dgl==0.7.0

Datasets

• Cora (available in data/)
• Pubmed (available in data/)
• ogbn-arxiv from OGB
• Reddit from http://snap.stanford.edu/graphsage/

Install graphattack:

cd GraphAttack
pip install -e .

Quick Start

See demo.ipynb

Reproduce results in the paper

run

python evaluate_guard.py

Cite

if you find this repo helpful, please cite our work:

@inproceedings{li2022guard,
  author       = {Jintang Li and
                  Jie Liao and
                  Ruofan Wu and
                  Liang Chen and
                  Zibin Zheng and
                  Jiawang Dan and
                  Changhua Meng and
                  Weiqiang Wang},
  title        = {{GUARD:} Graph Universal Adversarial Defense},
  booktitle    = {{CIKM}},
  pages        = {1198--1207},
  publisher    = {{ACM}},
  year         = {2023}
}