PROSPECT: Learn MLPs Robust against Graph Adversarial Structure Attacks

Table of Contents

• About
• Getting Started
• Usage
• Contributing

About

PROSPECT is the FIRST ONLINE graph distillation multi-layer perceptron (GD-MLP) framework PROSPECT, which learns GNNs and MLPs RObust againSt graPh advErsarial struCture atTacks.

In PROSPECT, an MLP and a GNN collaboratively learn by mutual distillation, as depicted by the following figure.

The highlights of PROSPECT have three aspects: 1) (provable) structure adversarial robustness, 2) adaptability to heterophilous graphs , and 3) the inference scalability. A rough comparison between PROSPECT and previous related works is shown by the table below, where 1-3 stars means low, medium, and high scores. Please refer to the paper for more details and discussion.

	Evasion structure robustness	Poi. structure robustness	Adaptability to heterophily	Inference scalability
Purification-based defenders	⭐️	⭐️ ⭐️	⭐️	⭐️
EvenNet	⭐️ ⭐️	⭐️ ⭐️	⭐️ ⭐️	⭐️
GLNN	⭐️ ⭐️	⭐️ ⭐️	⭐️ ⭐️	⭐️ ⭐️ ⭐️
PROSPECT	⭐️ ⭐️ ⭐️	⭐️ ⭐️ ⭐️	⭐️ ⭐️ ⭐️	⭐️ ⭐️ ⭐️

Getting_Started

These instructions will get you a copy of the project up and running on your local machine for development and testing purposes. We test the code with RTX4090 and CUDA11.8.

Prerequisites

The python packages used are list in requirements.txt

# --------- pytorch --------- #
torch==2.0.1
torchvision==0.9.0
lightning==2.0.6
torchmetrics==1.0.2

# --------- pytorch geometric--------- #
torch-sparse==0.6.15
torch-cluster==1.6.1
torch-scatter== 2.1.1
torch-geometric== 2.3.1
ogb==1.3.6

# --------- hydra --------- #
hydra-core==1.3.2
hydra-colorlog==1.2.0

# --------- loggers --------- #
wandb
tensorboard

# --------- others --------- #
jsonargparse==4.24.1    # cli parameter management
rootutils==1.0.7        # standardizing the project root setup

Installing

Just unzip the compressed atk_data.zip to atk_data folder, and run prospect_main.py under the project directory.

Usage

We provide part of experiment config files (under exp_configs) during the review period and the full config files for all datasets will be released as soon as the paper is accepted.

You can run the following commands in a terminal to get some results.

python prospect_main.py --config exp_configs/polblogs_metattack_0.15_clean=0_itr=0.0++gnn=sage_hid=64.yaml
python prospect_main.py --config exp_configs/cora_metattack_0.15_clean=0_itr=0.0++gnn=sage_hid=64.yaml
python prospect_main.py --config exp_configs/uai_metattack_0.15_clean=0_itr=0.0++gnn=sage_hid=64.yaml
python prospect_main.py --config exp_configs/texas_metattack_0.2_clean=0_itr=0.0++gnn=sage_hid=64.yaml

Contributing

As stated in the appendix of our paper, to facilitate future research on GD-MLPs, we have designed some LightningModules to eliminate boilerplate code so that researchers can focus on designing their distillation modules. These LightningModules support

1. offline, online, and hybrid distillation;
2. multiple teachers and/or multiple students;
3. plug-in (customized) feature and/or logit distillation strategy modules;
4. mini-batch sampling or full-batch training and distillation;
5. unified and friendly data interfaces for both clean and attacked graph data;
6. easy switch between transductive, inductive and semi-inductive evaluation settings;
7. easy switch between poisoning and evasion attack evaluation settings;
8. flexible experiment monitoring and management. Once the paper is accepted, we will release our preattacked datasets, library code, experiment code, and hyperparameters.

Now this repo is ready to implement many GD-MLP frameworks. We are welcome to the pull requests of your exciting distillation modules! 🍻

Credits

Thank PyTorch, PyG, and Lightning teams for their efforts to build research-friendly frameworks. Also thank @asheve for the inspiration from their template https://github.com/ashleve/lightning-hydra-template.

Citation

If you find this repository useful in your research, please consider giving a star ⭐ and a citation.

@inproceedings{deng2024prospect,
    author = {Deng, Bowen and Chen, Jialong and Hu, Yanming and Xu, Zhiyong and Chen, Chuan and Zhang, Tao},
    title = {PROSPECT: Learn MLPs on Graphs Robust against Adversarial Structure Attacks},
    year = {2024},
    isbn = {9798400704369},
    publisher = {Association for Computing Machinery},
    address = {New York, NY, USA},
    url = {https://doi.org/10.1145/3627673.3679857},
    doi = {10.1145/3627673.3679857},
    booktitle = {Proceedings of the 33rd ACM International Conference on Information and Knowledge Management},
    pages = {425–435},
    numpages = {11},
    keywords = {adversarial machine learning, graph heterophily, graph knowledge distillation, graph neural network},
    location = {Boise, ID, USA},
    series = {CIKM '24}
}