lizi-learner / MD-GNN

MD-GNN: A Lightweight Metric Defence Strategy Against Adversarial Attacks

MD-GNN

This repository contains the implementation of our paper：A Lightweight Metric Defence Strategy for Graph Neural Networks Against Poisoning Attacks accepted by ICICS2021.

Requirement

• matplotlib == 3.2.2

• torch_geometric == 1.5.0

• networkx == 2.4

• tqdm == 4.46.1

• torch == 1.5.0

• seaborn == 0.10.1

• scipy == 1.4.1

• numpy == 1.18.1

• openpyxl == 3.0.6

• deeprobust == 0.2.1

• scikit_learn==0.24.2

Run Test

• run test

  python test.py

• run MD-GCN

  # Supported Datasets：['cora', 'citeseer', 'cora_ml']
  # Supported Attacks：['meta', 'nettack', 'random']
  # Supported Metrics：['Cfs', 'Cfs1', 'Cfs2', 'Cfs3', 'Cfs4', 'Cs', 'Cs1', 'Jaccard1']
  device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
  
  data = Dataset(root='data/', name='cora', seed=15, require_mask=True)
  data.adj = sp.load_npz('adv_adj/cora_meta_adj_0.25.npz')
  output, time = CFS1(data, device, 'Cfs')

• run experiments

  from src.generate_experiment_data import *
  from src.generate_experiment_figure import *
  from src.save_excel import *
  
  dataset_list = ['cora', 'citeseer', 'cora_ml']
  attack_list = ['meta', 'nettack', 'random']
  defense_list = ['GCN', 'GCNSVD', 'Jaccard', 'ProGNN', 'RGCN', 'GAT', 'CFS', 'CfsGAT', 'CfsRGCN', 'GNNGuard', 'HGCN', 'CfsHGCN']
  root = 'data/'
  root1 = 'adv_adj/'
  save_dir = 'experiment/data/0/'
  
  threshold_list = [0.03, 0.05]
  a = [0.05, 0.1]
  lr_list = [0.001, 0.005, 0.01, 0.05, 0.1]
  
  # run experiments
  experiment1(dataset_list, attack_list,defense_list, root, root1, save_dir)
  experiment2(dataset_list, attack_list, threshold_list, root, root1, save_dir)
  experiment3(dataset_list, attack_list, root, root1, save_dir, a)
  experiment4(dataset_list, attack_list, defense_list, root, root1, save_dir, 'Cfs4')
  experiment5(dataset_list, attack_list, lr_list, root, root1, save_dir, 'Cfs', 10)
  
  # generate experimental results
  figure1()
  figure2()
  figure3()
  figure4()
  
  result()
  result_time()

• others

  run generate_adv.py # generate adversarial attacks

  run generate_clean.py # generate pre-processed graph

Project Structure

• MD-GNN
  • adv-adj
  • data
  • defense
  • experiment
  • src：source code
  • generate_adv.py
  • generate_clean.py
  • test.py
  • utils.py