urialon / SINGLE

Single-Node Attack For Fooling Graph Neural Networks

Single-Node Attack For Fooling Graph Neural Networks

This repository is the official implementation of Single-Node Attack For Fooling Graph Neural Networks.

Requirements

To install requirements:

pip install -r requirements.txt

To download the Twitter dataset:

wget https://www.dropbox.com/s/wmlfy463dqs07hu/twitter-dataset.tar.gz
tar -xvzf twitter-dataset.tar.gz
mv twitter-dataset/data/* GNNattack/datasets/twitter

Attacking

You can choose one of the 5 attacks as detailed in our paper:

1. SINGLE (AKA, NODE) SINGLE attack will produce a 2d matrix of SINGLE approaches such as (hops, GradChoice, Topology...) as a function of the available nets (GCN, GIN, GAT, SAGE)

2. EDGE EDGE attack will produce a 2d matrix of EDGE approaches such as (EdgeGrad, GlobalEdgeGrad, RANDOM) as a function of the available nets (GCN, GIN, GAT, SAGE)

3. NODE_LINF NODE_LINF attack will produce a 2d matrix of L_inf values {0.1-1.1} as a function of the available nets (GCN, GIN, GAT, SAGE), only for the basic SINGLE approach

4. DISTANCE DISTANCE attack will produce a 2d matrix of distance from the victim node as a function of the available nets (GCN, GIN, GAT, SAGE), only for the basic SINGLE approach

5. ADVERSARIAL ADVERSARIAL attack will produce a 1d vector of Ktest results for Ktrain=attEpochs, only for the basic SINGLE approach and for the GCN net

The available input arguments are:

• --attMode: Name of the attack Mode as described above

• --dataset: Name of the dataset, all caps

• --singleGNN: name of the wanted GNN (only in the case that you want results for ONE GNN)

• --num_layers: number of layers in the GNN

• --patience: the patience of the basic training (not the adversarial training)

• --attEpochs: number of attack epochs per victim node / number of Ktrain

• --lr: learn rate

• --l_inf: L_inf value, only for datasets that are represented in tf/idf (and not in many-hot-vec)

• --targeted: a bool flag that changes the attack to a targeted attack

• --distance (ONLY FOR THE DISTANCE ATTACK): the maximum distance

• --seed: a seed for reproducability