Multi-Target Constrained Graph Attack

A framework for training GNN models for fraud detection and evaluating their robustness against adversarial attacks.

Project Structure

project/
├─ data/
├─ models/
├─ results/
├─ src/
│  ├─ attack.py
│  ├─ config.py
│  ├─ dataloader.py
│  ├─ graph.py
│  ├─ models.py
│  ├─ predict.py
│  ├─ train.py
│  └─ utils.py
├─ config.yaml
├─ main.py
└─ README.md

Features

• Train GNN models (GCN, GAT, GATv2, SAGE, Chebyshev) for fraud detection
• Evaluate model performance on test data
• Execute Multi-Target Constrained Graph Attacks (MTCGA) on detected fraud nodes
• Analyze attack success rates and resource usage
• Configurable via config.yaml

Quick Start

1. Install dependencies

   uv pip install -r requirements.txt

2. Train models

   uv run python main.py train --models all --dataset mtcga

3. Evaluate models

   uv run python main.py evaluate --models GCN GAT --split test

4. Attack detected fraud nodes

   uv run python main.py attack --models Chebyshev --split test --limit 10

5. Analyze attack results

   uv run python main.py analyze --results results

Commands

• train - Train GNN models on fraud detection dataset
• evaluate - Evaluate trained models on specified data split
• attack - Execute adversarial attacks on detected fraud nodes
• analyze - Generate summary statistics from attack results

Configuration

Key parameters in config.yaml:

# Model training
lr: 0.005
epochs: 250
hidden_units: 256

# Attack parameters
p_evasion_threshold: 0.5
max_budget_prop: 0.4
max_transformations: 10
gas_penalty_coef: 0.1
value_penalty_coef: 0.01