readme.md

PGA

source code of Simple and Efficient Partial Graph Adversarial Attack: A New Perspective

Main Structure

• models: implementations of GNN models
• victims: training GNN models
  • configs: hyperparameters of models
  • models: saving checkpoints
• attackers: implementations of attack methods
• attack: perform attacks
  • configs: hyperparameters of attacks
  • perturbed_adjs: adversarial graphs generated by attackers

Running Steps

1. Training GNN models

> cd victims
> python train.py --model=gcn --dataset=cora

2. Perform attacks

> cd attack
> python gen_attack.py 

PGA attack

1. Training GNN models

> cd victims
> python train.py

2. Generate statistics of graphs, such as node degrees, classification margin and etc.

> cd analysis
> python gen_statistics.py --dataset=cora

3. Perform PGA attack

> cd attack
> python gen_attack.py --attack=pga --dataset=cora

Evaluate(evasion attack)

> python evasion_attack.py --victim=robust --dataset=cora
> python evasion_attack.py --victim=normal --dataset=cora

Evaluate(poisoning attack)

> python poison_attack.py --victim=gcn --dataset=cora
> python poison_attack.py --victim=gat --dataset=cora

requirements

• deeprobust
• torch_geometry
• torch_sparse
• torch_scatter