We actively support the following versions with security updates:

We take security vulnerabilities seriously. If you discover a security issue, please follow these steps:

Please do not report security vulnerabilities through public GitHub issues. This helps prevent malicious actors from exploiting the vulnerability before a fix is available.

Report the vulnerability privately through one of these methods:

• GitHub Security Advisories (Preferred):

  • Go to the Security tab
  • Click "Report a vulnerability"
  • Fill out the form with details

• Direct Email:

  • Contact the maintainer through GitHub
  • Use the subject line: [SECURITY] Brief description of vulnerability

Please include:

• Description of the vulnerability
• Steps to reproduce the issue
• Potential impact (e.g., data exposure, code execution)
• Affected versions
• Suggested fix (if you have one)
• Your contact information for follow-up

Example Report:

**Vulnerability Type:** Arbitrary Code Execution

**Description:**
The model loading function uses `torch.load()` without checking file
integrity, allowing malicious actors to embed arbitrary code in
pretrained weight files.

**Steps to Reproduce:**

1. Create malicious .pth file with embedded code
2. Load using `model.load_weights('malicious.pth')`
3. Arbitrary code executes during unpickling

**Impact:**
High - Could lead to remote code execution on user machines

**Affected Versions:**
All versions <= 1.0.0

**Suggested Fix:**
Use `torch.load(..., weights_only=True)` in PyTorch 2.0+
or verify checksums before loading

• Initial Response: Within 48 hours
• Status Update: Within 7 days
• Fix Timeline: Varies by severity
  • Critical: Emergency patch within 7 days
  • High: Patch within 30 days
  • Medium: Patch within 90 days
  • Low: Next regular release

• We follow coordinated disclosure
• Security advisories will be published after a fix is available
• You will be credited (unless you prefer anonymity)
• We aim for a 90-day disclosure window

# ✅ SAFE: Use weights_only for PyTorch 2.0+
model.load_state_dict(
    torch.load('model.pth', weights_only=True)
)

# ❌ UNSAFE: Default loading can execute arbitrary code
model.load_state_dict(
    torch.load('untrusted_model.pth')  # Dangerous!
)

# Always verify checksums for downloaded models
sha256sum pretrained_model.pth
# Compare with published checksum

Run untrusted models in isolated environments:

# Use Docker for isolation
docker run --rm -it \
    --gpus all \
    -v $(pwd):/workspace \
    deraining:latest \
    python scripts/inference.py --weights untrusted.pth

When processing user-uploaded images:

# Validate image before processing
def safe_load_image(path: str, max_size: int = 10 * 1024 * 1024):
    """Safely load and validate image.

    Args:
        path: Path to image file
        max_size: Maximum file size in bytes (default: 10MB)
    """
    # Check file size
    if os.path.getsize(path) > max_size:
        raise ValueError("Image too large")

    # Use safe image loading
    img = cv2.imread(path)
    if img is None:
        raise ValueError("Invalid image file")

    # Validate dimensions
    h, w = img.shape[:2]
    if h > 4096 or w > 4096:
        raise ValueError("Image dimensions too large")

    return img

Keep dependencies updated:

# Check for known vulnerabilities
pip install safety
safety check

# Update packages
pip install --upgrade -r requirements.txt

• Pickle Vulnerability: PyTorch model files (.pth) use pickle, which can execute arbitrary code
• Mitigation: Only load models from trusted sources, use weights_only=True in PyTorch 2.0+

• Large Images: Can cause memory exhaustion
• Mitigation: Validate image dimensions before processing

• Third-party libraries: May contain vulnerabilities
• Mitigation: Regular dependency updates, security scanning

• OOM Attacks: Maliciously large inputs can crash GPU
• Mitigation: Input size validation, memory monitoring

Security patches will be announced through:

• GitHub Security Advisories
• Release notes with [SECURITY] tag
• Updated CHANGELOG.md

Responsible disclosure researchers will be credited in:

• Security advisory
• Release notes
• CONTRIBUTORS.md (if desired)

For non-security questions about safety best practices:

• Open a GitHub Discussion
• Check existing security documentation

Thank you for helping keep this project secure! 🔒