Sparse but Strong: Crafting Adversarially Robust Graph Lottery Tickets

This is the official repository for Sparse but Strong: Crafting Adversarially Robust Graph Lottery Tickets which has been accepted at the NeurIPS 2023 GLFrontiers

Installation
Usage
Baseline Results

Installation

This project requires Python 3.8 and the following Python libraries installed:

torch==1.13.1
dgl-cu116
scikit-learn
networkx
scipy
matplotlib
pickle-mixin
argparse
numpy
warnings

To install these dependencies, navigate to your project's directory and run the following command:

pip install -r requirements.txt

Usage

You can test the performance of the trained adversarially robust GLTs by executing these commands on the terminal:

Evaluating the performance of ARGLTs for Cora and Citeseer Datasets attacked by PGD attack

python main_test.py --dataset cora --embedding-dim 1433 512 7 --attack_name pgd --ptb_rate 0.05
python main_test.py --dataset cora --embedding-dim 1433 512 7 --attack_name pgd --ptb_rate 0.1
python main_test.py --dataset cora --embedding-dim 1433 512 7 --attack_name pgd --ptb_rate 0.15
python main_test.py --dataset cora --embedding-dim 1433 512 7 --attack_name pgd --ptb_rate 0.2

python main_test.py --dataset citeseer --embedding-dim 3703 512 6 --attack_name pgd --ptb_rate 0.05
python main_test.py --dataset citeseer --embedding-dim 3703 512 6 --attack_name pgd --ptb_rate 0.1
python main_test.py --dataset citeseer --embedding-dim 3703 512 6 --attack_name pgd --ptb_rate 0.15
python main_test.py --dataset citeseer --embedding-dim 3703 512 6 --attack_name pgd --ptb_rate 0.2

Baseline Results

Here are the baseline accuracies for GCN on the Cora and Citeseer datasets under PGD attack for comparison:

GCN with PGD Attack on Cora Dataset

Perturbation Rate	Accuracy
0.05	78.81
0.1	78.46
0.15	77.93
0.2	77.99

GCN with PGD Attack on Citeseer Dataset

Perturbation Rate	Accuracy
0.05	73.28
0.1	73.96
0.15	74.62
0.2	73.52

You can also train the models by executing the following commands on the terminal:

Training ARGLT for Cora Dataset

 python main_train.py --dataset cora --embedding-dim 1433 512 7 --lr 0.008 --weight-decay 8e-5 --pruning_percent_wei 0.2 --pruning_percent_adj 0.05 --total_epoch 10 --s1 1e-2 --s2 1e-2 --init_soft_mask_type all_one --attack_name mettack --ptb_rate 0.05 --k 30 --alpha 1 --beta 0.01 --gamma 1 --alpha_fix_mask 1 --gamma_fix_mask 1

SubhajitDuttaChowdhury/ARGS