Unveiling the Threat of Fraud Gangs to Graph Neural Networks: Multi-Target Graph Injection Attacks against GNN-Based Fraud Detectors

This is the official code and data of the following paper:

Jinhyeok Choi, Heehyeon Kim, and Joyce Jiyoung Whang, Unveiling the Threat of Fraud Gangs to Graph Neural Networks: Multi-Target Graph Injection Attacks against GNN-Based Fraud Detectors, The 39th Annual AAAI Conference on Artificial Intelligence (AAAI), 2025

All codes are written by Jinhyeok Choi ([email protected]). When you use this code, please cite our paper.

@article{monti,
  author={Jinhyeok Choi and Heehyeon Kim and Joyce Jiyoung Whang},
  title={Unveiling the Threat of Fraud Gangs to Graph Neural Networks: Multi-Target Graph Injection Attacks against GNN-Based Fraud Detectors},
  year={2024},
  journal={arXiv preprint arXiv.2412.18370},
  doi={10.48550/arXiv.2412.18370}
}

Requirements

We used Python 3.8, Pytorch 1.13.1, and DGL 1.1.2 with cudatoolkit 11.7.

Usage

We conducted our experiments using GeForce RTX 2080 Ti, RTX 3090, or RTX A6000 with Intel(R) Xeon(R) Gold 6330 CPU @ 2.00GHz, running on Ubuntu 18.04.5 LTS.

Note: For experiments using GAGA, the initial pre-processing step may take 5-10 minutes before the actual training/inference begins for the first run.

Setup

Before running the commands below, please download the necessary pre-trained models and datasets:

1. Download best_models.zip and Datasets.zip from the following Google Drive link: https://drive.google.com/drive/folders/1E2FE8zDIS4XZA6Fxe5l3hTrXUnfhJ2gr?usp=sharing
2. Extract the downloaded files into their corresponding directories:
   • Extract best_models.zip into the ./best_models directory.
   • Extract Datasets.zip into the ./Datasets directory.

Reproducing Results

To reproduce the results of the best models, you can run the following command:

python run.py --config_path "./best_models/configs/GCN-GAGA-gossipcop_0.05-best_config.json" --inference

You can modify the config_path to experiment with different configurations and reproduce the other results.

Training from Scratch

You can run the following command to train MonTi from scratch with the best configuration:

python run.py --config_path "./best_models/configs/GCN-GAGA-gossipcop_0.05-best_config.json"

You can modify the config_path to experiment with different configurations and reproduce the other results.

Training logs and checkpoints will be saved in ./experimental_results.

Hyperparameters

The configuration files and checkpoints of the best models are in ./best_models.

The list of arguments of the configuration file:

{
	"exp_name": "Name of the experiment.",
	"model": "Attack model type to be used, e.g., 'MonTi'.",
	"surrogate_model": "Surrogate model type to be used, e.g., 'GCN'.",
	"victim_model": "Victim model type to be used, e.g., 'GAGA'.",
	"dataset": "Dataset to be used, e.g., 'gossipcop_0.05'.",
	"multi_target": "Whether the attack is multi-target or not.",
	"split": "Split number to be used.",
	"seed": "Random seed for the reproducibility of results.",
	"node_budget": "Node budget contorl paramter rho.",
	"edge_budget": "Edge budget contorl paramter xi.",
	
	"k_exp": "Hop number to extract neighbors from target set.",
	"masking": "Whether MonTi use the attack node masking or not.",
	"max_candidates": "Maximum number of candidates to consider during candidate selection (MonTi).",
	"num_layers": "Number of layers in the Transformer encoder.",
	"num_heads": "Number of heads in the Transformer encoder.",
	"emb_dim": "Dimension of each embedding.",
	"ff_dim": "Dimension of the feed-forward network in the Transformer encoder.",
	"dropout": "Dropout rate used in the model.",
	
	"initial_temp": "Initial temperature used in the temperature scheduling.",
	"min_temp": "Minimum temperature used in the temperature scheduling.",
	"temp_decay_rate": "The rate at which the temperature is decayed over training epochs.",
	"initial_eps": "Initial exploration parameter.",
	"min_eps": "Minimum exploration parameter.",
	"eps_decay_rate": "The rate at which the exploration parameter is decayed over training epochs.",
	
	"lr": "Learning rate for training.",
	"weight_decay": "Weight decay rate used for regularization during training.",
	"epochs": "Total number of training epochs.",
	"valid_epoch": "Number of epochs after which the validation is performed.",
	"patience": "Number of epochs to wait before early stopping if no progress on the validation set.",
	"batch_size": "Size of the batches used during training.",
	"batch_size_pre": "Size of the batches for pre-computing node representations.",
	
	"cuda_id": "CUDA device ID (GPU ID) to be used for training if available."
}

Description of each file and directory

./

• run.py: The main script to start the model training and evaluation.
• template.json: The template configuration file for multi-target attack experiments.
• template_single.json: The template configuration file for single-target attack experiments.

./best_models

• checkpoints: Pickle files containing the checkpoint of the trained model.
• configs: JSON files containing the best model configuration.

./Datasets

• Benchmark: Contains DGLGraph file, data splits, and target set assignment matrix for OGB-Prod and PubMed.
• GossipCop: Contains DGLGraph file, data splits, and target set assignment matrix for GossipCop-S.
• YelpChi_RTR: Contains DGLGraph file, data splits, and target set assignment matrix for YelpChi.

./models

• pretrained_models: Contains configurations and checkpoints of surrogate and victim models.
• attack_model.py: Contains the definition of MonTi.
• victim_model.py: Contains the definition of GCN and GAGA.

./modules

• data_handler.py: Manages data loading and preprocessing.
• experiment_handler.py: Handles the setup and execution of experiments.
• result_manager.py: Manages the logging and saving of experiment results.
• schedulers.py: Contains schedulers for early stopping and exponential decaying.

./utils

• constants.py: Defines constants used across the project.
• utils.py: Helper functions used throughout the project.