Defense Against Knowledge Poisoning Attack on GraphRAG

🧠 Overview

This repository contains the code and experiments for a defense framework against knowledge poisoning attacks in GraphRAG-style multi-hop QA systems. Our method, HoG-GRAG (Hop-wise Guard for GraphRAG), improves robustness by decomposing multi-hop questions into ordered subqueries, detecting poisoning-induced inconsistencies during hop-wise execution, and repairing corrupted retrieved subgraphs through targeted pruning and minimal evidence recovery.

We evaluate this framework on multi-hop question answering using:

• Datasets

  • MuSiQue
  • HotpotQA

• GraphRAG Pipelines

  • Microsoft GraphRAG
  • LightRAG

📂 Repository Structure

This repository is organized around the two core components of Auto-Immune GraphRAG — detection and repair — with supporting modules for evaluation.

HoG-GRAG/
│        
├── prompts/ 
│   │   └── Question Paraphrasing.md
│   │   └── Response Evaluation.md
├── src/
│   ├── detection.py               
│   ├── repairer.py             
│   ├── trace_analysis.py               
│   ├── evaluation.py
├── baselines/
│   │   └── Query_Paraphrasing.py
│   │   └── Perplexity_based.py
├── requirements.txt             
└── README.md

📄 Citation

If you use this methodology in your research, please cite:

Havva Alizadeh Noughabi, Fattane Zarrinkalam, Ali Dehghantanha, Defense Against Knowledge Poisoning Attack on GraphRAG, Accepted at the Annual Meeting of the Association for Computational Linguistics (ACL 2026)