README.md

Adversarial Graph Benchmark (AGB)

Towards Practical and Fair Evaluation of Adversarial Graph Neural Networks

---

Benchmark • Robustness • Leaderboards • Research

Live Website: https://agb.chartalist.org

What is AGB?

AGB is a benchmark platform designed to provide standardized, fair, and reproducible evaluation of adversarial attacks and defenses on Graph Neural Networks (GNNs).

Inspired by the GOttack benchmark framework, AGB enables researchers to compare methods under consistent settings across multiple datasets, attack models, and defense strategies.

---

Features

Leaderboards Compare attack and defense performance using unified evaluation protocols. Benchmark Datasets Homophilic, heterophilic, and large-scale graph benchmarks. Attack Evaluation Fair comparison of poisoning and evasion attacks.

Defense Analysis Evaluate robustness of defended GNN models. Reproducibility Standardized experimental settings. Research Platform Designed for the graph learning community.

---

Benchmark at a Glance

Experiments	Attacks	Defenses	Datasets
437,000+	7	8	6

---

Supported Datasets

Homophilic	Heterophilic	Large Scale
CORA	CHAMELEON	OGB-ARXIV
CITESEER	SQUIRREL
PUBMED

---

Website Pages

Page	Description
Home	Benchmark overview
Leaderboard	Benchmark rankings
Datasets	Dataset collection
AGB Challenge	Challenge tracks
Paper	Benchmark publication
GitHub	Source code

---

Quick Start

git clone https://github.com/FDataLab/AGB.github.io.git cd AGB.github.io

cd AGB-website

npm install

npm run dev

Open:

http://localhost:4000

---

Contributors

Developed by researchers and contributors from:

• University of Central Florida
• Virginia Tech

---

Support

If you find AGB useful:

Star this repository

Share with the graph learning community

Use it in your research

---

Built for the Adversarial Graph Learning Community

AGB • Fair Evaluation • Reproducible Research

Built for the Graph Learning and Adversarial Machine Learning Research Community