Data Poisoning Attack on RAG Apps

📚 FAISS-based Text Search

A simple pipeline to chunk text, generate embeddings, and perform semantic search using faiss for fast retrieval.

👾 Data Poisoning Attack on RAG

Create a malicious .txt file and inject it into the vectordb to poison the knowledge and the RAG App.

⚠️ Results

Base RAG Response	Data Poison Attack: RAG Response

---

📦 Setup

1. clone the repository

   git clone https://github.com/dhruvXcode247/Data-Poisoning-Attack-On-LLM-RAG.git
   cd Data-Poisoning-Attack-on-LLM-RAG

2. create a virtual environment (optional)

   python -m venv venv
   source venv/bin/activate  # windows: venv\Scripts\activate

3. install dependencies

   pip install -r requirements.txt

---

📊 Usage

1. Get Text Files:
   Run create_dataset.ipynb to create a dataset from scratch.

   export GROQ_API_KEY="<groq-api-key>"

2. Create VectorDB:
   Run create_vectorstore.ipynb to generate the HuggingFace Embeddings (intfloat/multilingual-e5-small) and store in FAISS Vector Store.

3. Poison Data Injection:
   Run poison_data.py script which adds malicious payload to single .txt file, and creates relevant vector stores for retreival.

   python poison_data.py

4. Run Streamlit App
   Run the app.py python script.

   streamlit start app.py

---

🚀 Features

• load and process .txt files from a directory
• use HuggingFace embeddings for vector representation
• store and retrieve chunks with faiss
• persistent storage with faiss index + json

---

📜 todo

• create streamlit app for user interaction
• create poisoned .txt file
• add poisoned .txt into vector store
• include vector store selection (base/poison) in streamlit app
• test data poisoning vulnerability

---

🛠️ Tech stack

• LangChain
• FAISS
• Groq API

Data-Poisoning-Attack-On-LLM-RAG

This repository implements a data poisoning attack done on a LLM-RAG application hosted on Streamlit using FAISS.