Membership Inference Attacks to GNN for Graph Classification

The source code for ICDM2021 paper: "Adapting Membership Inference Attacks to GNN for Graph Classification: Approaches and Implications". The full version of the paper can be found in https://arxiv.org/abs/2110.08760

Installation

If you meet the version mismatch error for Lasagne library, please use following command to upgrade Lasagne library.

pip install --upgrade https://github.com/Lasagne/Lasagne/archive/master.zip

Usage

Step 1: Train Target and Shadow Model

To train the victim GCN model, use the following command.

For common graph dataset (DD, PROTEINS_full, ENZYMES):

bash run_TUs_target_shadow_training.sh --number 10 --start_epoch 100 --dataset DD

./run_TUs_target_shadow_training.sh --number 10 --start_epoch 100 --dataset DD --pretrain_epochs 50 --use_pretrained

./run_TUs_target_shadow_training.sh --number 1 --start_epoch 100 --dataset DD --pretrain_epochs 100 --use_pretrained

For graph converted via SuperPixel (CIFAR10, MNIST)

bash run_SPs_target_shadow_training.sh --number 10 --start_epoch 100 --dataset MNIST

bash run_SPs_target_shadow_training.sh --number 1 --start_epoch 100 --dataset CIFAR10

--number this is the number of repeated model training

--start_epoch this is the minimum number of interactions to train the model

--dataset this is the dataset name of model training

For detailed code execution, you can refer to 'main_SPs_graph_classification.py' and 'main_TUs_graph_classification.py' in the ./code folder.

Step 2: Train Attack Model Inferring Graph Sample Membership

bash run_transfer_attach.sh --number 15

--number this is the number of repeated attack

For detailed code execution, you can refer to 'transfer_based_attack.py'.

srun python3 -m venv ~/mia_gnn
source ~/mia_gnn_env/bin/activate

pip install dgl==2.0.0 -f https://data.dgl.ai/wheels/cu118/repo.html