Awesome

Single-Node Attack For Fooling Graph Neural Networks

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

Requirements

This project is based on PyTorch 1.6.0 and the PyTorch Geometric library.

First, install PyTorch from the official website: https://pytorch.org/. Then install PyTorch Geometric: https://pytorch-geometric.readthedocs.io/en/latest/notes/installation.html (PyTorch Geometric must be installed according to the instructions there). Eventually, run the following to verify that all dependencies are satisfied:

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/* ./datasets/twitter

Attacking

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

1. SINGLE NODE 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, SGC, Robust GCN...)

2. SINGLE EDGE attack will produce a 2d matrix of EDGE approaches such as (SINGLE, GradChoice...) as a function of the basic available nets (GCN, GIN, GAT, SAGE, SGC)

3. NODE_LINF attack will produce a 2d matrix of L_inf values as a function of the available nets, only for the basic SINGLE approach

4. NODE_L0 attack will produce a 2d matrix of L_0 values as a function of the available nets, only for the basic SINGLE approach

5. DISTANCE attack will produce a 2d matrix of distance from the victim node as a function of the available nets, only for the basic SINGLE approach

6. ADVERSARIAL 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, SGC...), for a model which is trained adversarialy on the basic SINGLE approach

7. MULTIPLE attack will produce a 2d matrix of the number of attackers as a function of the available nets, only for the basic SINGLE approach

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: the learning rate

• --l_inf: the L_inf value, the limit on the maximal change of an attribute that is used for the attack. Available only for datasets that are not represented in as a many-hot-vec or a one-hot-vec

• --l_0: the L_0 value, the limit on the ratio of attributes used for the attack

• --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

Note: Every combination of attack mode and GNN is available, except for the combination of Edge attacks+Robust GNNs