MAF (LLM Model Application Firewall)

A comprehensive security framework for LLM applications, providing content filtering, compliance checking, and prompt safety analysis. MAF acts as a protective layer between your application and LLM services, ensuring safe and compliant interactions.

Core Functions

1. Content Security

   • PII (Personal Identifiable Information) Detection and Masking
   • Prompt Injection Detection
   • Input/Output Sanitization

2. Compliance Checking

   • Islamic Content Compliance
   • Cultural Sensitivity Analysis
   • Content Policy Enforcement

3. LLM Safety

   • Prompt Safety Verification
   • Response Validation
   • Model Output Filtering

Features

• Prompt Injection Detection
• User Input Sanitization
• Return Value Validation
• Semantic Analysis
• Log Management
• Web-based Management Interface

Setup

1. Create a virtual environment:

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

2. Install dependencies:

pip install -r requirements.txt

3. Set up environment variables: Create a .env file with the following variables:

HUGGINGFACE_TOKEN=your_token_here
SECRET_KEY=your_secret_key_here

4. Run the application:

uvicorn app.main:app --reload

Project Structure

maf/
├── app/
│   ├── main.py              # FastAPI application entry point
│   ├── core/
│   │   ├── config.py        # Configuration settings
│   │   ├── security.py      # Security utilities
│   │   └── logging.py       # Logging configuration
│   ├── models/
│   │   └── schemas.py       # Pydantic models
│   ├── services/
│   │   ├── prompt_checker.py    # Prompt injection detection
│   │   ├── sanitizer.py         # Input sanitization
│   │   └── validator.py         # Return value validation
│   └── api/
│       └── endpoints.py     # API endpoints
├── tests/                   # Test files
├── logs/                    # Application logs
└── requirements.txt         # Project dependencies

API Documentation

Once the application is running, visit:

• Swagger UI: http://localhost:8000/docs
• ReDoc: http://localhost:8000/redoc

Security Features

1. Prompt Injection Detection

   • Analyzes user input for potential injection attacks
   • Uses semantic analysis to identify malicious patterns

2. Input Sanitization

   • Automatically removes sensitive information
   • Logs sanitization actions for audit purposes

3. Return Value Validation

   • Ensures output complies with security rules
   • Prevents sensitive data leakage

4. Logging

   • Comprehensive logging of all security-related events
   • File-based storage for audit trails

Running the Application

Docker (single image, frontend + backend)

docker compose up --build -d

The React frontend and FastAPI backend are packaged into one image. FastAPI serves the built frontend from /app/static with SPA fallback, and the API under /api/v1.

.env needs HUGGINGFACE_TOKEN with read access to the private model repo yutang718/evyd-defender (EVYD Defender; version-less repo, current weights tagged v4); the other guard models are public. Model weights are cached in the model_cache volume after the first start. To run a locally trained checkpoint instead, set MAF_GUARD_MODEL_PATH=/app/trained_models/<dir> (the ./models directory is mounted there).

• Web UI + API: http://localhost:8080
• OpenAPI docs: http://localhost:8080/docs

Local development

1. Start the Backend Server:

./server.sh
# or
uvicorn app.main:app --host 0.0.0.0 --port 8000 --reload

2. Start the Frontend Application:

cd frontend
streamlit run main.py

Frontend Features

1. Prompt Safety Checking

   • Real-time prompt analysis
   • Safety score visualization
   • Detailed threat analysis
   • Model selection options

2. PII Detection Interface

   • Interactive text analysis
   • PII highlighting and masking
   • Rule configuration management
   • Detection statistics

3. Islamic Compliance System

   • Comparative analysis view
   • Rule management interface
   • Multi-language support
   • Compliance testing tools

Frontend Access

• Main application: http://localhost:8501
• Prompt checking: http://localhost:8501/prompt_checking
• PII filtering: http://localhost:8501/pii_filtering
• Islamic rules: http://localhost:8501/islamic_rules