SECURITY.md

VREyeSAM - Model Security & Protection Guide

🔒 Overview

VREyeSAM is protected with multiple security layers to prevent model weight extraction and ensure safe deployment.

Security Measures Implemented

1. Model Weight Protection

• ✅ Model weights are loaded at startup and never exposed to the client
• ✅ Weights are managed in model_server.py using a singleton pattern
• ✅ Checkpoint paths are resolved internally and never sent to the frontend

2. File System Isolation

• ✅ Checkpoint files have restricted permissions (600)
• ✅ Only the inference API is exposed to users
• ✅ Raw file access is blocked

3. API-Only Architecture

• ✅ No direct model file downloads
• ✅ Only prediction results are returned to users
• ✅ Model internals stay hidden

Deployment to Hugging Face Spaces

Prerequisites

1. HuggingFace account with Spaces access
2. Model weights in private HuggingFace repository
3. Docker setup for containerized deployment

Step 1: Create Private Model Repository

# Clone your model repo (if not already done)
# Ensure checkpoints are NOT committed to git
# Add to .gitignore if needed

Step 2: Deploy to HF Spaces

1. Go to Hugging Face Spaces
2. Click "Create new Space"
3. Fill in details:
   • Space name: vreyesam
   • License: MIT
   • SDK: Docker
   • Visibility: Public (only code, not weights)
4. After creation, upload your Dockerfile and code files

Step 3: Authentication for Model Downloads

For accessing private model weights during Docker build:

1. Create HuggingFace token: https://huggingface.co/settings/tokens
2. Set in Spaces environment (Settings → Secrets with HF_TOKEN)
3. OR use direct URL with token (not recommended, keep private)

Step 4: Verify Security

Before deployment:

# Check what files will be uploaded
git status
git ls-files | grep -E '\.(pt|pth|torch|bin)$'

# Should output: (nothing - no weights!)

Security Checklist

• Model weights are in .gitignore
• Checkpoint paths are not hardcoded in code
• Only model_server.py handles weight loading
• Docker build uses secure downloads
• .env files are in .gitignore
• Frontend cannot access file paths
• API only exposes prediction results

Best Practices

✅ DO:

• Keep model weights private and download during deployment
• Use environment variables for configuration
• Only expose prediction API endpoints
• Log errors without exposing paths
• Use Hugging Face tokens securely in Spaces secrets

❌ DON'T:

• Commit model weights to git
• Hardcode checkpoint paths in code
• Expose debug routes that show model structure
• Log full file paths to users
• Include weights in Docker layers visible to users

Troubleshooting

Issue: "Model weights not found"

1. Verify .gitignore contains checkpoint paths
2. Check Dockerfile correctly downloads from HuggingFace
3. Ensure HF_TOKEN is set in Spaces secrets

Issue: "File path exposed in error"

1. Update model_server.py to not show paths
2. Generic error messages only: "Model initialization failed"
3. Check logs don't contain sensitive details

Advanced Security

Optional: Encrypt Weights

# In model_server.py
from cryptography.fernet import Fernet
encrypted_weights = Fernet(key).encrypt(state_dict)

Optional: Disable Direct File Access

# Set file permissions
chmod 600 segment-anything-2/checkpoints/*
# Only the app process can read them

Support

For security questions or issues:

• Check the GitHub Issues
• Contact: geetanjalisharma546@gmail.com

---

Last Updated: March 2025 Security Level: High Protection ✅