LoadingByte / are-gnn-defenses-robust

Adaptive evaluation reveals that most examined adversarial defenses for GNNs show no or only marginal improvement in robustness. (NeurIPS 2022)

Are Defenses for Graph Neural Networks Robust?

This is the Python 3.9 code we have used to study the adversarial robustness of various GNN defenses in our NeurIPS 2022 paper "Are Defenses for Graph Neural Networks Robust?".

Project Page • Paper (OpenReview) • Video (YouTube)

A cursory reading of the literature suggests that we have made a lot of progress in designing effective adversarial defenses for Graph Neural Networks (GNNs). Yet, the standard methodology has a serious flaw – virtually all of the defenses are evaluated against non-adaptive attacks leading to overly optimistic robustness estimates. We perform a thorough robustness analysis of 7 of the most popular defenses spanning the entire spectrum of strategies, i.e., aimed at improving the graph, the architecture, or the training. The results are sobering – most defenses show no or only marginal improvement compared to an undefended baseline. We advocate using custom adaptive attacks as a gold standard and we outline the lessons we learned from successfully designing such attacks. Moreover, our diverse collection of perturbed graphs forms a (black-box) unit test offering a first glance at a model's robustness.

Please cite our paper if you refer to our results or use the method or code in your own work:

@inproceedings{mujkanovic2022_are_defenses_for_gnns_robust,
    title = {Are Defenses for Graph Neural Networks Robust?},
    author = {Mujkanovic, Felix and Geisler, Simon and G\"unnemann, Stephan and Bojchevski, Aleksandar},
    booktitle = {Neural Information Processing Systems, {NeurIPS}},
    year = {2022}
}

Quick Start

To introduce yourself to the various parts of the code, take a look at the demo notebook in notebooks/demo.ipynb.

If you are looking for the black-box robustness unit test, you can find a self-explanatory demo script in unit_test/sketch.py.

Setup

To install all required dependencies, run the script requirements-install.sh. If you are using a non-Unix machine, run these commands manually instead:

$ pip install -r requirements-torch.txt -f "https://download.pytorch.org/whl/<PLATFORM>/torch_stable.html"
$ pip install -r requirements.txt

Code Organization

• data/ – The versions of the Cora ML and Citeseer datasets used by us.
• gb/ – The main codebase. Contains both experiment code and analysis tools.
• ex_config/ – Configurations to reproduce our experiments. See the next section for details.
• ex_*.py – Entry points for the experiments configured in ex_config/.
• notebooks/ – The Jupyter notebooks mentioned in this readme file.
• unit_test/ – The best perturbations per model that we transfer to other defenses in Fig. 7, as well as a self-explanatory script sketching how to use them to perform a black-box robustness unit test on any defense.

Reproducing the Experiments

We use Sacred and run the experiments on our internal Slurm cluster and use MongoDB to keep track of the experiments via the SEML library. The folder ex_config/ contains the SEML configurations to reproduce every single attack we have run.

Depending on the SEML and Sacred settings, resources like data might not be available at runtime. To remedy this, manually make a copy of this source folder available on each worker machine and point the variable FALLBACK_SRC_PATH in gb/util.py to it.

The evasion_*.yml experiments – which actually also include meta attacks for technical reasons – need to be run first. Note, however, that some of these attacks require the output of other attacks, which they use as their initialization. Currently, the corresponding run IDs are substituted with TODO in the ex_config/ files. Hence, the experiments must be executed in order, and you need to manually insert the run IDs once they are available. You can obtain the run IDs using the MongoDB queries referenced in the configuration files directly above the TODO substitutes.

As soon as all "evasion" experiments are finished, you need to use the MongoDB queries cited in trans_poisoning.yml to obtain and insert all evasion experiment run IDs. Then, start off those experiments as well.

Finally, run the cross_model_*.yml experiments to transfer perturbations between models.

Next, utilize the notebook data_extraction.ipynb to download all experiment results from MongoDB and extract useful information from them in the form of long strings. These strings you then paste into paper_plots.ipynb at the appropriate locations to reproduce the exact plots we present in our paper and the appendix.

Note, however, that – while possible – reproducing all experiments takes a considerable amount of time (see also §5 in the paper).