Add GNN code and configurations

.gitignore

@@ -18,3 +18,7 @@ dist/
 
 #virtual environments folder
 .venv
+temp_results/
+robustness_results/
+*.pth
+*.pt

README.md

@@ -1,118 +1,18 @@
-# PyGIP
+# PyGIP - GNN Ownership Verification Module
 
-[![PyPI - Version](https://img.shields.io/pypi/v/PyGIP)](https://pypi.org/project/PyGIP)
-[![Build Status](https://img.shields.io/github/actions/workflow/status/LabRAI/PyGIP/docs.yml)](https://github.com/LabRAI/PyGIP/actions)
-[![License](https://img.shields.io/github/license/LabRAI/PyGIP.svg)](https://github.com/LabRAI/PyGIP/blob/main/LICENSE)
-[![PyPI - Downloads](https://img.shields.io/pypi/dm/pygip)](https://github.com/LabRAI/PyGIP)
-[![Issues](https://img.shields.io/github/issues/LabRAI/PyGIP)](https://github.com/LabRAI/PyGIP)
-[![Pull Requests](https://img.shields.io/github/issues-pr/LabRAI/PyGIP)](https://github.com/LabRAI/PyGIP)
-[![Stars](https://img.shields.io/github/stars/LabRAI/PyGIP)](https://github.com/LabRAI/PyGIP)
-[![GitHub forks](https://img.shields.io/github/forks/LabRAI/PyGIP)](https://github.com/LabRAI/PyGIP)
+This repository contains the integration of **Graph Neural Network (GNN) ownership verification** experiments into the PyGIP framework. It provides modular and extensible implementations for attacks and defenses on GNNs, following the guidelines of the PyGIP framework.
 
-PyGIP is a Python library designed for experimenting with graph-based model extraction attacks and defenses. It provides
-a modular framework to implement and test attack and defense strategies on graph datasets.
+---
 
-## How to Cite
+## 📋 Overview
 
-If you find it useful, please considering cite the following work:
+This module allows users to:  
 
-```bibtex
-@article{li2025intellectual,
-  title={Intellectual Property in Graph-Based Machine Learning as a Service: Attacks and Defenses},
-  author={Li, Lincan and Shen, Bolin and Zhao, Chenxi and Sun, Yuxiang and Zhao, Kaixiang and Pan, Shirui and Dong, Yushun},
-  journal={arXiv preprint arXiv:2508.19641},
-  year={2025}
-}
-```
+- Evaluate ownership verification on GNN models (GCN, GAT, GraphSAGE).  
+- Run experiments under **inductive** and **transductive** settings.   
+- Easily extend the framework with new datasets, attacks, or defenses.
 
+> **Note:** Large model weights (`benign_model.pth`) and result folders (`temp_results/`, `robustness_results/`) are excluded to keep the repository clean.  
 
-## Installation
+---
 
-PyGIP supports both CPU and GPU environments. Make sure you have Python installed (version >= 3.8, <3.13).
-
-### Base Installation
-
-First, install the core package:
-
-```bash
-pip install PyGIP
-```
-
-This will install PyGIP with minimal dependencies.
-
-### CPU Version
-
-```bash
-pip install "PyGIP[torch,dgl]" \
-  --index-url https://download.pytorch.org/whl/cpu \
-  --extra-index-url https://pypi.org/simple \
-  -f https://data.dgl.ai/wheels/repo.html
-```
-
-### GPU Version (CUDA 12.1)
-
-```bash
-pip install "PyGIP[torch,dgl]" \
-  --index-url https://download.pytorch.org/whl/cu121 \
-  --extra-index-url https://pypi.org/simple \
-  -f https://data.dgl.ai/wheels/torch-2.3/cu121/repo.html
-```
-
-## Quick Start
-
-Here’s a simple example to launch a Model Extraction Attack using PyGIP:
-
-```python
-from datasets import Cora
-from models.attack import ModelExtractionAttack0
-
-# Load the Cora dataset
-dataset = Cora()
-
-# Initialize the attack with a sampling ratio of 0.25
-mea = ModelExtractionAttack0(dataset, 0.25)
-
-# Execute the attack
-mea.attack()
-```
-
-This code loads the Cora dataset, initializes a basic model extraction attack (`ModelExtractionAttack0`), and runs the
-attack with a specified sampling ratio.
-
-And a simple example to run a Defense method against Model Extraction Attack:
-
-```python
-from datasets import Cora
-from models.defense import RandomWM
-
-# Load the Cora dataset
-dataset = Cora()
-
-# Initialize the attack with a sampling ratio of 0.25
-med = RandomWM(dataset, 0.25)
-
-# Execute the defense
-med.defend()
-```
-
-which runs the Random Watermarking Graph to defend against MEA.
-
-If you want to use cuda, please set environment variable:
-
-```shell
-export PYGIP_DEVICE=cuda:0
-```
-
-## Implementation & Contributors Guideline
-
-Refer to [Implementation Guideline](.github/IMPLEMENTATION.md)
-
-Refer to [Contributors Guideline](.github/CONTRIBUTING.md)
-
-## License
-
-[BSD 2-Clause License](LICENSE)
-
-## Contact
-
-For questions or contributions, please contact blshen@fsu.edu.

config/global_cfg.yaml

@@ -0,0 +1,13 @@
+target_model: gat
+target_hidden_dims: [224, 128]
+dataset: Cora
+train_setting: ""
+test_setting: 1
+embedding_dim: 128
+train_process: "train"
+test_process: "test"
+n_run: 3
+
+train_save_root: ../temp_results/diff/model_states/
+test_save_root: ../temp_results/diff/model_states/
+res_path: ../temp_results/diff/res/

config/test_setting1.yaml

@@ -0,0 +1,4 @@
+model_arches: ["gat", "gcn", "sage"]
+layer_dims: [96,  160, 224, 288, 352]
+num_hidden_layers: [2]
+num_model_per_arch: 10

config/test_setting2.yaml

@@ -0,0 +1,4 @@
+model_arches: ["gat", "gcn", "sage"]
+layer_dims: [128, 192, 256, 320, 384]
+num_hidden_layers: [1, 3]
+num_model_per_arch: 10

config/test_setting3.yaml

@@ -0,0 +1,4 @@
+model_arches: ["gin", 'sgc']
+layer_dims: [96,  160, 224, 288, 352]
+num_hidden_layers: [2]
+num_model_per_arch: 15

config/test_setting4.yaml

@@ -0,0 +1,4 @@
+model_arches: ["gin", 'sgc']
+layer_dims: [128, 192, 256, 320, 384]
+num_hidden_layers: [1, 3]
+num_model_per_arch: 15

config/train_setting.yaml

@@ -0,0 +1,4 @@
+model_arches: ["gat", "gcn", "sage"]
+layer_dims: [96,  160, 224, 288, 352]
+num_hidden_layers: [2]
+num_model_per_arch: 20 # per arch