🛡️ LLM-Security-Firewall

🚨 Semantic Prompt Injection Firewall for Large Language Models

---

📌 Overview

LLM-Security-Firewall is a lightweight semantic firewall designed to protect Large Language Models (LLMs) against:

• 🚫 Prompt Injection Attacks
• 🔓 Jailbreak Attempts
• 🧠 Role-play Exploits
• ⚠️ Instruction Override Attacks
• 🔐 System Prompt Leakage

The system introduces a Decoupled Security Architecture that intercepts and analyzes prompts before they ever reach the LLM.

Unlike traditional keyword filtering systems, this project uses:

• 🧠 Semantic Vector Embeddings
• ⚡ XGBoost Threat Classification
• 📊 Real-time Probability Analysis
• 🛡️ Local Security Inference

This allows the firewall to detect malicious intent even when attacks are obfuscated or rephrased.

---

🏗️ System Architecture

The firewall operates using a 4-Phase Security Pipeline:

Phase	Description
📥 Input & Ingestion	User submits prompt through Streamlit UI
🧠 Semantic Analysis	MiniLM converts prompt into embeddings
🚦 Logic Gate	XGBoost classifies malicious probability
☁️ Secure Resolution	Only verified prompts reach the LLM

---

🔐 Threat Detection Logic

The security core performs:

Prompt → Embedding → XGBoost → Threat Score → Decision

⚙️ Detection Pipeline

Step 1 — Semantic Vectorization

The prompt is transformed into a 384-dimensional semantic vector using:

all-MiniLM-L6-v2

Step 2 — Classification

The embedding is analyzed using an XGBoost classifier.

Step 3 — Security Decision

Threat Score	Action
> 0.5	🚫 BLOCK Prompt
≤ 0.5	✅ Forward to LLM

---

💬 User Interaction Flow

The user experience is designed for:

• ⚡ Real-time detection
• 📊 Instant feedback
• 🛡️ Transparent security status
• ☁️ Secure API routing

Malicious prompts are intercepted locally before any external transmission occurs.

---

✨ Features

🛡️ Security

• Prompt Injection Detection
• Jailbreak Prevention
• Semantic Threat Analysis
• Probability-based Threat Scoring
• Decoupled LLM Firewall

⚡ Performance

• Lightweight ML inference
• Low latency (<50ms)
• CPU-friendly deployment
• Real-time classification

📊 Interface

• Interactive Streamlit Dashboard
• Plotly Threat Meter
• Session Threat Logs
• Safe/Blocked Visual Alerts

☁️ API Integration

• OpenRouter Integration
• Secure Request Routing
• Conditional API Access
• External LLM Protection Layer

---

🧠 Machine Learning Pipeline

Component	Technology
Embedding Model	Sentence Transformers
Semantic Encoder	all-MiniLM-L6-v2
Classifier	XGBoost
Frontend	Streamlit
Backend	FastAPI
Visualization	Plotly
API Integration	OpenRouter
Programming Language	Python

---

📈 Model Performance

Metric	Score
🎯 Accuracy	85.15%
📈 ROC-AUC	0.8989
⚖️ Weighted F1-Score	83.85%
🛡️ Precision (Malicious)	86.20%
🚨 Recall (Malicious)	81.50%

---

📂 Project Structure

LLM-Security-Firewall/
│
├── streamlit_app.py
├── api_server.py
├── config.py
├── requirements.txt
├── README.md
├── .gitignore
├── .env.example
├── LICENSE
│
├── models/
│   └── xgb_embed_model.json
│
├── notebook/
│   └── PromptInjectionSecurityFilter.ipynb
│
├── assets/
│   ├── architecture-overview.png
│   ├── technical-workflow.png
│   ├── security-decision-flow.png
│   └── chat-flow-diagram.png
│
└── docs/
    ├── Project_Report.pdf
    └── PSF_Presentation.pptx

---

🚀 Installation

1️⃣ Clone Repository

git clone https://github.com/YOUR_USERNAME/LLM-Security-Firewall.git

cd LLM-Security-Firewall

---

2️⃣ Install Dependencies

pip install -r requirements.txt

---

3️⃣ Configure Environment Variables

Create a .env file:

OPENROUTER_API_KEY=your_api_key_here

---

▶️ Running the Application

🔥 Launch Streamlit Frontend

streamlit run streamlit_app.py

---

⚡ Launch FastAPI Backend

uvicorn api_server:app --reload

---

🔄 Security Workflow

User Prompt
     ↓
Semantic Vectorization
     ↓
XGBoost Threat Classification
     ↓
Threat Score Evaluation
     ↓
┌───────────────┬────────────────┐
│ Malicious     │ Safe           │
│ Prompt        │ Prompt         │
├───────────────┼────────────────┤
│ BLOCK         │ Forward to LLM │
└───────────────┴────────────────┘

---

📚 Research Contributions

This project demonstrates that:

• Lightweight machine learning models can effectively secure LLMs
• Semantic embeddings outperform traditional keyword filters
• Decoupled architectures reduce inference costs
• Real-time LLM firewalls can operate on standard consumer hardware

---

🔮 Future Improvements

• 🧠 FAISS vector database integration
• 🌍 Multilingual attack detection
• 🖼️ Multimodal prompt injection defense
• 💬 Multi-turn conversational memory analysis
• ⚡ Adaptive threshold optimization
• ☁️ Cloud-native deployment support

---

📄 Documentation

Detailed research and implementation documents are available inside:

docs/

Including:

• 📘 Research Report
• 📊 Project Presentation
• 🧪 Experimental Results
• 🧠 ML Workflow Analysis

---

⭐ Support

If you found this project useful:

• ⭐ Star the repository
• 🍴 Fork the project
• 🛡️ Contribute to LLM security research

---

📜 License

This project is licensed under the MIT License.

---

🛡️ Secure the Future of AI

Building Defensive Intelligence for Large Language Models