harry24k/gnn-meta-attack

PyTorch implementation of Adversarial Attacks on Graph Neural Networks via Meta Learning.

GNN-Meta-Attack

This repository is a pytorch implementation of Adversarial Attacks on Graph Neural Networks via Meta Learning.

It is modified from following githubs :

https://github.com/ChandlerBang/pytorch-gnn-meta-attack (Main Reference, Pytorch)

https://github.com/danielzuegner/gnn-meta-attack (Official, Tensorflow)

https://github.com/Kaushalya/gnn-meta-attack-pytorch (Pytorch)

Requirements

• python 3.6
• torch 1.2
• torchvision 0.5
• higher
• numpy
• scipy
• matpotlib
• seaborn

Usage

See help (-h flag) for detailed parameter list of each script before executing the code.

Training

To train the model(s) in the paper, run this command:

# cora_ml
python main.py --train True --hidden 16 --data-name cora_ml --save-path sample.pth

Evaluation

To evaluate the model(s) saved locally, run this command:

# cora_ml
python main.py --train False --hidden 6 --data-name cora_ml --save-path sample.pth

Generate Poisoned Data with Meta Attack

Here is how to generate poisoned data :

# cora_ml
python poison.py --hidden 6 --lambda_ 0.5 --train-iters 15 --perturb-rate 0.05 --save-path sample.pth --data-name cora_ml

(+) Approximated Meta Attack is not implemented.

Results

The accuracy on each dataset :

(+) Unless it's mentioned, all values are set to default. (+) For each method, lambdas are 0(Self), 0.5(Both) and 1(Train).

cora_ml

Description	Perturb Rate	Accuracy(Dropped)	Reported	Data Name
Clean	0.00	85.80%	83.40%	cora_ml.npz
Self	0.05	80.43%(5.37%p)	75.50%(7.90%p)	cora_ml_self_5.npz
Both	0.05	81.42%(4.38%p)	-	cora_ml_both_5.npz
Train	0.05	82.52%(3.28%p)	78.00%(5.40%p)	cora_ml_train_5.npz
Self	0.20	58.01%(27.79%p)	-	cora_ml_self_20.npz
Both	0.20	67.73%(18.07%p)	-	cora_ml_both_20.npz
Train	0.20	80.03%(5.77%p)	-	cora_ml_train_20.npz

(+) Due to GPU Limitation, I used train-iters=15 when generating poisoned data. Thus, the accuracies of meta-attack likely to be higher than reported ones.

citeseer

Description	Perturb Rate	Accuracy(Dropped)	Reported	Data Name
Clean	0.00	70.41%	72.50%	citeseer.npz
Self	0.05	64.35%(6.06%p)	65.40%(7.10%p)	citeseer_self_5.npz
Both	0.05	61.08%(9.33%p)	-	citeseer_both_5.npz
Train	0.05	68.98%(1.43%p)	69.70%(2.80%p)	citeseer_train_5.npz

(+) Due to GPU Limitation, I used train-iters=15 when generating poisoned data. Thus, the accuracies of meta-attack likely to be higher than reported ones.

polblogs

Description	Perturb Rate	Accuracy(Dropped)	Reported	Data Name
Clean	0.00	93.18%	93.60%	polblogs.npz
Self	0.05	74.36%(18.82%p)	77.50%(16.10%p)	polblogs_self_5.npz
Both	0.05	78.27%(14.91%p)	-	polblogs_both_5.npz
Train	0.05	86.09%(7.09%p)	83.70%(9.90%p)	polblogs_train_5.npz

(+) Due to GPU Limitation, I used train-iters=15 when generating poisoned data. Thus, the accuracies of meta-attack alikely to be higher than reported ones.

Visualization

Here is a visualization of changes of edges with Lightning.

Please see Visulaize.ipynb to reproduce these images.

cora_ml	cora_ml_both_5

(+) Statistics

• Deleted Edges: 0
• Created Edges: 798
• Number of Nodes
  • Effected Train Nodes: 282
  • Effected Test Nodes: 202