TaintSentinel: Path-Level Randomness Vulnerability Detection for Smart Contracts

Overview

TaintSentinel is a novel dual-stream Graph Neural Network (GNN) approach designed to detect bad randomness vulnerabilities in Ethereum smart contracts. By combining global contract analysis with path-specific taint tracking, TaintSentinel achieves state-of-the-art performance in identifying weak entropy sources that can be exploited by malicious actors.

Key Features

• Dual-Stream Architecture: Combines global contract context with path-level analysis
• Gated Fusion Mechanism: Adaptive integration of global and local features
• Path Risk Assessment (PRA): Fine-grained risk classification for taint propagation paths
• High Accuracy: 87.5% F1-Score on balanced datasets

Performance

Metric	Balanced Dataset	Imbalanced Dataset
F1-Score	0.875	0.611
Precision	0.921	0.537
Recall	0.833	0.707
AUC-ROC	0.951	0.940
PRA	0.970	0.920

Repository Structure

• dataset/ - Smart contract dataset
  • contracts/ - All contract files
    • vulnerable/ - 231 vulnerable contracts
    • safe/ - 4,256 safe contracts
  • dataset_metadata.json - Dataset statistics
• preprocessing/ - 5-step preprocessing pipeline
  • complete_parser1.py - AST extraction
  • source2.py - Source detection
  • sink3.py - Sink detection
  • ModularSemanticGraph.py - Graph construction
  • TaintAnalysis.py - Taint path analysis
• model/ - Model implementation
  • Sentinel_1.py - Dataset loader
  • Sentinel_2.py - GNN models
  • Sentinel_3.py - Training pipeline
  • Sentinel_4.py - Enhanced with PRA
• requirements.txt - Dependencies

Quick Start

Prerequisites

• Python 3.8+
• PyTorch 1.12+
• PyTorch Geometric
• NumPy
• Pandas
• Scikit-learn
• Pathlib (built-in Python 3.4+)

Installation

git clone https://github.com/HadisRe/TaintSentinel.git

cd TaintSentinel

pip install -r requirements.txt

Dataset

The TaintSentinel dataset contains 4,487 smart contracts (231 vulnerable, 4,256 safe) organized in two batches. Dataset is included in the repository under dataset/contracts/.

Running the Preprocessing Pipeline

Execute the preprocessing steps in order:

cd preprocessing

python complete_parser1.py

python source2.py

python sink3.py

python ModularSemanticGraph.py

python TaintAnalysis.py

Training the Model

cd ../model

python Sentinel_1.py # Dataset loader testing

python Sentinel_2.py # Model architecture testing

python Sentinel_3.py # Full training pipeline

python Sentinel_4.py # Training with Path Risk Accuracy