This repository contains the source code for the paper, "Illuminati: Towards Explaining Graph Neural Networks for Cybersecurity Analysis" by Haoyu He, Yuede Ji and H. Howie Huang.
[Paper] [EuroS&P 2022]
We provide the implementation of public datasets for the purpose of reproduction.
We use Pytorch Geometric for implementation. Please follow the instruction to install Pytorch Geometric.
The datasets can be loaded from Pytorch Geometric. The datasets we use in the paper are:
["BBBP", "Mutagenicity", "BA-2motifs"] for graph classification.
["Cora", "Citeseer"] for node classification.
We modify the dataset loader from DIG.
TASK as graph for graph classification, node for node classification.
We use configs.json to control this project. Specifically,
mode - the choice of explanation methods {0: GNNExplainer or Illuminati, 1: PGM-Explainer, 2: PGExplainer}
node - whether to estimate node importance scores, i.e., GNNExplainer or Illuminati
synchronize - synchronized attribute mask learning
agg1 & agg2 - aggregation functions for node importance scores [mean, min, max, etc.]
sample - the number of samples for PGM-Explainer
$ python train_TASK.py
We save the model with the best validation result while training. We also provide early stop.
$ python explain_TASK.py
This is to explain the whole dataset. We save node importance scores for each graph as a single ".csv" file. The time complexity can be evaluated here.
$ python evaluate_TASK.py
Here, we provide Essentialness Percentage (EP) and probability reduction for explained subgraphs and remaining subgraphs.
@INPROCEEDINGS{illuminati22,
author={He, Haoyu and Ji, Yuede and Huang, H. Howie},
booktitle={2022 IEEE 7th European Symposium on Security and Privacy (EuroS\&P)},
title={Illuminati: Towards Explaining Graph Neural Networks for Cybersecurity Analysis},
year={2022},
volume={},
number={},
pages={74-89},
doi={10.1109/EuroSP53844.2022.00013}}