GNNCert: Deterministic Certification of Graph Neural Networks against Adversarial Perturbations

An official PyTorch implementation of "GNNCert: Deterministic Certification of Graph Neural Networks against Adversarial Perturbations" (ICLR 2024).

This code is based on the official GIN implementation: https://github.com/weihua916/powerful-gnns.

Installation

First, unzip the dataset.zip file.

cd GraphGuard
unzip dataset.zip

Then, create a new environment and install dependencies. We use Python 3.8 and PyTorch 1.10.0+cu113.

pip install -r requirements.txt

Note: In addition to the datasets in dataset.zip, you can also use the datasets from PyG. To use the PyG datasets, install the corresponding version of PyG according to https://pytorch-geometric.readthedocs.io/en/latest/install/installation.html.

Experments

To run GNNCert with GIN, use the following commands:

• For feature division:

bash jobs/feature_division.sh
bash jobs/feature_division-test.sh

• For structure division:

bash jobs/structure_division.sh
bash jobs/structure_division-test.sh

• For feature-structure jointly division:

bash jobs/method_structure_feature.sh
bash jobs/method_structure_feature-test.sh

You can also try different architectures, such as GCN and GAT:

• For GCN:

bash jobs/GCN.sh
bash jobs/GCN-test.sh

• For GAT:

bash jobs/GAT.sh
bash jobs/GAT-test.sh