huyoboy / graph-universal-attack

Graph Universal Adversarial Attack (GUA)

Usage

• PyTorch 0.4 or 0.5
• Python 2.7 or 3.6
• networkx, scipy, sklearn, numpy, pickle

Train the attack model

Example: python generate_perturbation.py --dataset cora --radius 4

dataset: the network dataset you are going to attack
radius: the radius of the l2 Norm Projection

The verision of jupyter notebook is also supported as: universal_attack.ipynb

Evaluate the test ASR

After finishing the training of the GUA, we then evaluate the test asr over the test nodes

Example: python eval_baseline.py --dataset cora --radius 4 --evaluate_mode universal

dataset: the network dataset you are going to attack
radius: the radius of the l2 Norm Projection evaluate_mode has five values:

• "universal": graph universal attack
• "limitted_attack": random attack, a prescribed number of anchor nodes are randomly sampled
• "global_random": global random attack
• "victim_attack": victim attack
• "universal_delete": randomly delete a part of nodes from the trained anchor nodes, to find the trade-off
• "advanced_victim_attack": a prescribed number of anchor nodes are composed of the nodes with the highest confidence from the victim class
• "advanced_limitted_attack": advanced random attack, a prescribed number of anchor nodes are randomly sampled from the top 10% nodes with the highest degrees

The perturbation results trained by GUA when radius = 4, for each dataset: Cora, Citeseer and Pol.Blogs are also provided in "GUA/perturbation_results", which can be used directly for testing

The verision of jupyter notebook is also supported as: evaluate.ipynb

You can also validate the transferability on other embedding methods:
Node2vec: python node2vec/evaluate_n2v --dataset cora
DeepWalk: python deepwalk/evaluate_deepwalk --dataset cora
pyGAT: python pyGAT/evaluate_GAT --dataset cora