KG-FakeBench: Generating and Detecting Knowledge-Graph-Grounded Fake Information with LLMs

KG-FakeBench is a large-scale benchmark for evaluating large language models (LLMs) on AI-generated misinformation under controlled factual deviations.

The benchmark leverages knowledge graphs (KGs) to generate misinformation that is factually incorrect yet semantically plausible, and introduces a KG-consistent evidence framework for structured, evidence-based detection.

📄 Paper: coming soon

---

🔑 Key Contributions

• KG-Grounded Benchmark
  A large-scale dataset generated via structured KG deviations, ensuring fine-grained control over factual plausibility and transparent provenance.

• KG-Consistent Evidence Detection
  A structured pipeline that extracts (s, r, o) triples and uses them as external evidence for LLM-based verification.

• Comprehensive Evaluation
  Evaluation across standard, CoT, and KG-grounded prompting shows that external evidence improves detection reliability and reveals model behavior and bias.

---

📊 Dataset

• 28,900 synthetic samples
  • 14,450 high-plausibility
  • 14,450 low-plausibility
• 1,239 real samples

The dataset is derived from WikiGraphs, ensuring structured and verifiable factual grounding.

📂 See data/ for details.

---

🧠 Methodology

KG-FakeBench consists of two main components:

🔹 1. KG-Driven Fake Information Generation

• Extract reference triples ⟨s, r, o⟩ from KG
• Generate fake triples ⟨s, r, o′⟩ with controlled plausibility
• Use LLMs to synthesize natural language misinformation

🔹 2. KG-Consistent Evidence Detection

• Tokenize and normalize input statements
• Retrieve candidate entities via similarity matching
• Ground statements into KG triples
• Use triple-based prompting for factual verification

---

📂 Repository Structure

KG-FakeBench/
├── data/                                   # Dataset (fake + real samples)
├── code/
│   ├── KG-Driven Fake Information Generation/   # Generation pipeline
│   ├── KG-Consistent Evidence Detection/        # Detection pipeline
├── README.md

---

⚙️ Code

The implementation is divided into:

• KG-Driven Fake Information Generation
• KG-Consistent Evidence Detection

📂 See code/ for details.

---

⚠️ Disclaimer

This dataset contains synthetic misinformation generated for research purposes only.
It is intended to support the development of robust and trustworthy detection systems.

---

🔗 Resources

• 📦 Dataset: data/
• ⚙️ Code: code/

---

🤝 Contributing

Contributions are welcome. Please open an issue or submit a pull request.