model-sentinel 0.2.0

A security verification tool for AI model scripts - Detects and verifies changes in Python files of AI models

🛡️ Model Sentinel

A security verification tool for model scripts - Detects and verifies changes in Python files of AI models.

Features

• Hugging Face Hub Model Verification: Detect changes in Python files of remote models
• Local Model Verification: Detect changes in model files in local directories
• Hash-based Verification: Verify file integrity using hashes
• Interactive Approval: Review and approve content of changed files
• GUI Support: Intuitive web-based GUI interface

Installation

Basic Version (CLI only)

pip install model-sentinel

GUI Version

pip install "model-sentinel[gui]"

Usage

CLI Usage

# Show help and usage instructions
model-sentinel

# Verify Hugging Face model
model-sentinel --hf ryomo/malicious-code-test

# Verify local model
model-sentinel --local ./my-model-directory

# List all verified models
model-sentinel --list-verified

# Delete all verification data
model-sentinel --delete

GUI Usage

Note: GUI commands require the GUI version to be installed.

model-sentinel --gui --hf ryomo/malicious-code-test
model-sentinel --gui --local ./my-model-directory

Python Script Usage

from model_sentinel import verify_hf_model, verify_local_model

# Verify Hugging Face model
result = verify_hf_model("ryomo/malicious-code-test")  # Returns True if verified, False otherwise

# Verify local model
result = verify_local_model("./my-model-directory")  # Returns True if verified, False otherwise

# Verify with GUI mode
result = verify_hf_model("ryomo/malicious-code-test", gui=True)  # GUI window will open

Verification Process

1. Hash Comparison: Calculate hash of entire model or directory and compare with previous verification
2. File Verification: If changes detected, check individual Python files
3. Content Display: Show content of changed files (pager in CLI, web interface in GUI)
4. User Approval: Only approve if user confirms content is trustworthy
5. Directory Update: Save file content and metadata to .model-sentinel/ directory structure

Verification Data Directory

Verification data is stored in a structured .model-sentinel/ directory:

.model-sentinel/
├── registry.json           # Global registry of verified models
├── local/                  # Local models
│   └── {model_name}_{hash}/
│       ├── metadata.json   # Model metadata and file info
│       └── files/          # Individual file content
└── hf/                     # HuggingFace models
    └── {org}/{model}@{revision}/
        ├── metadata.json
        └── files/

Example metadata.json:

{
  "model_hash": "abc123...",
  "last_verified": "2025-07-28T10:30:00Z",
  "files": {
    "modeling.py": {
      "hash": "def456...",
      "size": 1024,
      "verified_at": "2025-07-28T10:30:00Z"
    }
  }
}

Development

For development and contributing to this project:

# Clone and setup
git clone https://github.com/ryomo/model-sentinel.git
cd model-sentinel

# Install dependencies
uv sync

# Run from source (for testing)
uv run model-sentinel  # Show help
uv run model-sentinel --hf ryomo/malicious-code-test
uv run model-sentinel --local ./my-model-directory
uv run model-sentinel --gui --hf ryomo/malicious-code-test

Testing

This project uses Python's built-in unittest for testing.

Running Tests

Run all tests:

uv run python -m unittest discover tests -v

Run specific test module:

uv run python -m unittest tests.test_verify.test_verify -v
uv run python -m unittest tests.test_target.test_base -v
uv run python -m unittest tests.test_cli -v

Test Coverage

Generate coverage reports:

# Run tests with coverage
uv run python -m coverage run -m unittest discover tests

# Generate coverage report
uv run python -m coverage report --include="src/*"

# Generate HTML coverage report
uv run python -m coverage html --include="src/*"
# Open htmlcov/index.html in browser

Technical Specifications

• Python: 3.12+
• Package Manager: uv
• GUI Framework: Gradio 5.x
• Hash Algorithm: SHA-256
• Supported Files: Python files (.py)

License

This project is licensed under the MIT License.

Contributing

Pull requests and issue reports are welcome.