This repository contains the official implementation of:

S. Geisler, D. Zügner, and S. Günnemann. Reliable Graph Neural Networks via Robust Aggregation. Neural Information Processing Systems, NeurIPS, 2020

See also: Project Page - Arxiv - Google Colab Notebook

The main idea is to substitute the message passing aggregation in a Graph Neural Network (GNN)

with robust location estimators for improved robustness w.r.t. adversarial modifications of the graph structure.

In Figure 1 of our paper, we give an exemplary plot for Nettack that clearly shows that strong adversarially added edges are resulting in a concentrated region of outliers. This is exactly the case where robust aggregations are particularity strong.

We show that in combination with personalized page rank (aka GDC - Graph Diffusion Convolution) our method (aka Soft Medoid GDC) outperforms all baselines and tested state of the art adversarial defenses:

Consider citing our work via:

@inproceedings{geisler2020robustaggregation,
  title =      {Reliable Graph Neural Networks via Robust Aggregation},
  author =     {Simon Geisler, Daniel Z{\"{u}}gner and Stephan G{\"{u}}nnemann},
  booktitle =  {Neural Information Processing Systems, {NeurIPS}},
  year =       {2020},
}

Execute

conda install pytorch==1.6.0 torchvision torchaudio cudatoolkit=10.1 -c pytorch
pip install -r requirements.txt
pip install .

pip install ./kernels
conda install gmpy2 statsmodels
pip install ./sparse_smoothing

for setting the project up. Run for the results on empirical robustness (takes about 4 minutes with a GPU):

python script_evaluate_empirical_robustness.py

For the certified robustness via randomized smoothing use:

python script_evaluate_certified_robustness.py

For simplicity we recommend to install PyTorch a priori via anaconda:

conda install pytorch==1.6.0 torchvision torchaudio cudatoolkit=10.1 -c pytorch

We used Python 3.7.6 and CUDA 10.1. We provide custom CUDA kernels that are fairly simple implementations for a row-wise topk and row-wise weighted median on a sparse matrix.

Due to custom CUDA kernels, you must be able to compile via nvcc. Conda handles the c++ compiler etc. You also must have installed the CUDA toolkit and should select the matching CUDA version for your environment. Note that PyTorch Geometric and PyTorch have some version-dependent restriction regarding the supported CUDA versions. See also Build PyTorch from source which captures the requirements for building custom extensions.

If you simply want to use the CPU, you can do not need to bothera bout CUDa and can go on with the installation. Later on you must add --kwargs '{"device": "cpu"}' while executing the script_*.py files.

Thereafter we can install the actual module via (alternatively use python install .):

pip install -r requirements.txt
pip install .

By default the requirements are installed with very restrictive versioning since we did not test any other configuration. If you have version conflicts, you can also build without version restrictions via omitting the command pip install -r requirements.txt (not tested).

In case you want to use the GPU, you also need to fulfill the requirements for compiling a custom C++/CUDA extension for PyTorch - usually satisfied by default voa the conda command above.

You can either build the kernels a priori with

pip install ./kernels

or PyTorch will try to compile the kernels at runtime.

If you want to run the randomized smoothing experiments you need to install the respective module:

conda install gmpy2 statsmodels
pip install ./sparse_smoothing

In case the installation of gmpy fails please check out their installation guide.

To (unit) test the robust mean functions, you can run (make sure pytest is on your path):

    pytest tests

We also provde the reqwuirements we used during development via:

pip install -r requirements-dev.txt

Note: you can skip this section as we provide pretrained models

For the training and evaluation code we decided to provide SEML/Sacred experiments which make it very easy to run the same code from the command line or on your cluster.

The training for all the pretrained models is bundled in:

python script_train.py --kwargs '{"artifact_dir": "cache"}'

To train a model on cora_ml for evaluating the empirical robustness (from then on it will be also used for evaluation) e.g. run:

python experiment_train.py with "dataset=cora_ml" "seed=0" "model_params={\"label\": \"Soft Medoid GDC (T=1.0)\", \"model\": \"RGNN\", \"do_cache_adj_prep\": True, \"n_filters\": 64, \"dropout\": 0.5, \"mean\": \"soft_k_medoid\", \"mean_kwargs\": {\"k\": 64, \"temperature\": 1.0}, \"svd_params\": None, \"jaccard_params\": None, \"gdc_params\": {\"alpha\": 0.15, \"k\": 64}}" "artifact_dir=cache" "binary_attr=False"

With binary attributes (for randomized smoothing) use:

python experiment_train.py with "dataset=cora_ml" "seed=0" "model_params={\"label\": \"Soft Medoid GDC (T=1.0)\", \"model\": \"RGNN\", \"do_cache_adj_prep\": True, \"n_filters\": 64, \"dropout\": 0.5, \"mean\": \"soft_k_medoid\", \"mean_kwargs\": {\"k\": 64, \"temperature\": 1.0}, \"svd_params\": None, \"jaccard_params\": None, \"gdc_params\": {\"alpha\": 0.15, \"k\": 64}}" "artifact_dir=cache" "binary_attr=True"

For evaluation, we execute all locally stored (pretrained) models.

Similarly to training, we provide a script that runs the attacks for different seeds for all pretrained models:

python script_evaluate_empirical_robustness.py

This will print the following table:

For Cora ML and Citeseer run

python script_evaluate_certified_robustness.py

This command results in:

Note that we use a different setup than in the paper. For example, here we updated to PyTorch 1.6.0 and use the most recent sparse smoothing code. This is the main reason why the numbers are slightly different than in our paper.

This code is licensed under MIT. In you want to contribute, feel free to open a pull request.