Official implementation of "Inference Attacks Against Graph Neural Networks" (USENIX Security 2022)
Overview of the Code
The code entry is main.py, it will invoke the experimental classes in 'exp/' folder to conduct different experiments. For example, if you run the code from command line
python main.py --attack 'graph_reconstruction'
Notice that the arguments are optional, you can specify the default values in the main.py file.
Code Structure
- main.py: The entry of the whole project, including the configurations of the logger, the parameters used in the experiments.
- config.py: Define some constants used in the experiments, including various data paths.
- exp/: This folder contains the main procedure of different experiments, where each experiment correspond to one class. They will invoke functions implemented in other 'lib_' folders.
- lib_dataset/: This folder contains the implementations for data pre-precessing. We will store the preprocessed data in pickle format to speed up the main experiments.
- lib_gnn_model/: This folder contains different gnn models.
- lib_plot/: This folder contains code for plotting
- temp_data/: This folder is used to store the raw data, the processed data and some temporary data. I do not include this folder in the github, you can create the required folders following the path specified in config.py.
Directory Structure under temp_data
.
├── attack_data
│ ├── graph_reconstruct
│ ├── property_infer
│ ├── property_infer_basic
│ └── subgraph_infer
├── defense_data
│ ├── property_infer
│ └── subgraph_infer
├── gae_model
│ ├── AIDS_diff_pool
│ ├── AIDS_mean_pool
│ ├── AIDS_mincut_pool
│ └── fine_tune
├── model
│ ├── model_AIDS
│ ├── para_AIDS
├── original_dataset
│ ├── AIDS
├── split
│ └── 20
└── target_model
└── 20
Citation
@inproceedings{zhang22usenix,
author = {Zhikun Zhang and Min Chen and Michael Backes and Yun Shen and Yang Zhang},
title = {{Inference Attacks Against Graph Neural Networks}},
booktitle = {USENIX Security Symposium (USENIX Security)},
year = {2022}
}