Certifying Robust Graph Classification under Orthogonal Gromov-Wasserstein Threats

Preparation

• create virtual env (restrict to python=3.7, which is compatible with CPLEX 12.10) conda create -n gdro python=3.7

• install pytorch / pytorch-geometric conda install pytorch=1.8.0 torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia pip install torch-scatter torch-sparse torch-cluster torch-spline-conv -f https://pytorch-geometric.com/whl/torch-1.8.0+cu111.html pip install torch-geometric pot

• install supplimentary libraries conda install matplotlib joblib networkx numba conda install gurobi -c gurobi pip install qpsolvers nsopy

• install the package as develop mode python setup.py develop

• (optional) install additional solver if available. conda install docplex cplex -c imbdecisionoptimization conda install mosek -c mosek conda install cvxopt -c conda-forge

dataset

dataset	# graphs	# lables	# features	ave. edge	min edge	max edge	avg. node	min node	max node
MUTAG	188	2	7	38	20	66	17.5	10	28
PTC_MR	344	2	18	25.0	2	142	13.0	2	64
COX2	467	2	38	86.0	68	118	41.0	32	56
BZR	405	2	56	74.0	26	120	35.0	13	57

Experiments

We provided a comprehensive notebook demo.ipynb to show the idea of

• tractable bounds of FGW
• convex extension of FGW
• certification and attack for task of graph classification

Other files includes:

• demo_train.py: build the model for certificate and attack.
• demo_spla.py: preprocess to generate the linear mapping matrix $\mathcal{A}$.
• demo_certify.py: complete setup for experiments of robust certifications.
• demo_attack.py: complete setup for experiments of attack.

License

The project is under MIT license.

NOTE:

scipy==1.7.3 in case the changes in source code made differences.