MIA-GCL

This is the implementation for our paper "GCL-LLeak: Link Membership Inference Attacks against Graph Contrastive Learning", which has been accepted by PoPETs 2024.

Datasets

• The datasets of Cora, Citeseer, Amazon-computer, Amazon-photo can be download with the package of pytroch Geometric,

• The datasets of ENZYMES, COX2, Google+ are provided in the file "data"

• The dataset of Facebook can be download here: https://snap.stanford.edu/data/ego-Facebook.html

• The datasets of Cora, Citeseer with different density are provided in the file "data/density"

• The datasets of Facebook - Ego and Google+ with different homophily are provided in the file "data/homophily"

GCL models (target model)

The original implemenations of GCL models we used in the paper can be found here:

• GRACE: https://github.com/CRIPAC-DIG/GRACE

• MVGRL: https://github.com/kavehhassani/mvgrl

• GCA: https://github.com/CRIPAC-DIG/GCA

• CCA-SSG: https://github.com/hengruizhang98/CCA-SSG

• MERIT: https://github.com/merit-gem/merit

Thanks for the authors providing the implementations.

Requirements

We tested the implementations with the following reqirements:

• Python 3.8

• dgl-cuda11.3 0.9.0

• torch 1.10.0+cu113

• torch-geometric 2.0.3

• torch-scatter 2.0.9

• torch-sparse 0.6.15

Attacks against GRACE

python GRACE-mia-white.py

Attacks against MAGRL

python MVGRL-mia-white.py

Attacks against GCA

python python train-mia-white.py --device cuda:0 --dataset Cora --param local:cora.json --drop_scheme degree

Attacks against CCA-SSG

python main-cora-mia-white.py

Attacks against MERIT

python train-cora-mia.py

Evaluate the defense mechanisms

For DP-SGD

python GRACE-mia-white-dpsgd-defense.py

python MVGRL-mia-white-dpsgd-defense.py

python main-cora-mia-white-dpsgd-defense.py

For Noisy embedding

python GRACE-mia-white-lap-defense.py

python MVGRL-mia-white-lap-defense.py

python main-cora-mia-white-lap-defense.py