Defending Graph Convolutional Networks against Dynamic Graph Perturbations via Bayesian Self-supervision

Authors: Jun Zhuang, Mohammad Al Hasan.

Paper:

Accepted by AAAI 2022

Abstract:

In recent years, plentiful evidence illustrates that Graph Convolutional Networks (GCNs) achieve extraordinary accomplishments on the node classification task. However, GCNs may be vulnerable to adversarial attacks on label-scarce dynamic graphs. Many existing works aim to strengthen the robustness of GCNs; for instance, adversarial training is used to shield GCNs against malicious perturbations. However, these works fail on dynamic graphs for which label scarcity is a pressing issue. To overcome label scarcity, self-training attempts to iteratively assign pseudo-labels to highly confident unlabeled nodes but such attempts may suffer serious degradation under dynamic graph perturbations. In this paper, we generalize noisy supervision as a kind of self-supervised learning method and then propose a novel Bayesian self-supervision model, namely GraphSS, to address the issue. Extensive experiments demonstrate that GraphSS can not only affirmatively alert the perturbations on dynamic graphs but also effectively recover the prediction of a node classifier when the graph is under such perturbations. These two advantages prove to be generalized over three classic GCNs across five public graph datasets.

Dataset:

cora, citeseer, pubmed, amazoncobuy, coauthor

Getting Started:

Prerequisites

Linux or macOS
CPU or NVIDIA GPU + CUDA CuDNN
Python 3
pytorch, dgl, numpy, scipy, sklearn, numba

Clone this repo

git clone https://github.com/junzhuang-code/GraphSS.git
cd GraphSS/graphSS

Install dependencies

For pip users, please type the command: pip install -r requirements.txt
For Conda users, you may create a new Conda environment using: conda env create -f environment.yml

Directories

graphSS:

1. pretrain.py: Train the node classifier on the train graph
2. attack.py: Implement adversarial attacks
3. main.py: Infer labels by GraphSS
4. models_GCN.py/models_SS.py: model scripts
5. load_data.py: Load data script
6. utils.py: Utils modules

data:

1. attacker_data: a directory that stores attacked graph and GCN's weights
2. noisy_label: a directory that stores labels

Runs

1. Train the node classifier on the train graph
   python [script_name] -data_name -model_name -NUM_EPOCHS -GPU -NOISY_RATIO -is_trainable
   e.g. python pretrain.py cora GCN 200 0 0.1 true

2. Apply adversarial attacks and generate predicted labels
   python [script_name] -data_name -model_name -TARGET_CLASS -attack_type -NUM_PERT -SAMPLE_RATE -GPU -NOISY_RATIO -is_attack
   e.g. python attack.py cora GCN -1 "lf" 2 0.2 0 0.1 true

3. Inference and evaluation
   python [script_name] -data_name -model_name -is_attacked -TARGET_CLASS -attack_type -SAMPLE_RATE -GPU -is_inferable -ssup_type -NUM_RETRAIN -NUM_INFER -WARMUP_STEP -Alert -Dynamic_Phi
   e.g. python main.py cora GCN true -1 "lf" 0.1 0 true "pseudo" 60 100 40 "" true

4. Visualization with Tensorboard
   Under the directory of "GraphSS", run: Tensorboard --logdir=./runs/Logs_SS --port=8999

Source of Competing Methods:

AdvTrain: When Does Self-Supervision Help Graph Convolutional Networks? [Code]
GNN-Jaccard: Adversarial Examples on Graph Data: Deep Insights into Attack and Defense [Code]
GNN-SVD: All You Need Is Low (Rank): Defending Against Adversarial Attacks on Graphs [Code]
RGCN: Robust Graph Convolutional Networks Against Adversarial Attacks [Code]
GRAND: Graph Random Neural Network for Semi-Supervised Learning on Graphs [Code]
ProGNN: Graph Structure Learning for Robust Graph Neural Networks [Code]
NRGNN: NRGNN: Learning a Label Noise Resistant Graph Neural Network on Sparsely and Noisily Labeled Graphs [Code]

Cite

Please cite our paper if you think this repo is helpful.

@article{zhuang2022defending,
  author={Zhuang, Jun and Al Hasan, Mohammad},
  year={2022},
  month={Jun.},
  title={Defending Graph Convolutional Networks against Dynamic Graph Perturbations via Bayesian Self-Supervision},
  journal={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={36},
  number={4},
  pages={4405-4413},
  DOI={10.1609/aaai.v36i4.20362},
  url={https://ojs.aaai.org/index.php/AAAI/article/view/20362}
}