Skip to content

Security: Lemniscate-world/Neural

docs/security.md

Security in Neural

Overview

Neural uses Flask and Flask-SocketIO for its API and WebSocket endpoints, with security measures to protect data and access.

Authentication

  • HTTP Basic Auth: The /trace API endpoint and WebSocket connections require basic authentication.
  • Configuration: Credentials are stored in config.yaml:
    auth:
      username: "admin"
      password: "your_secure_password"
    
    

Hacky Mode

  • Purpose: Inspired by "HackingNeuralNetworks," --hacky mode analyzes gradient leakage and simulates adversarial inputs to detect security vulnerabilities.
  • Usage: neural debug my_model.neural --hacky
  • Features:
    • Gradient Leakage: Checks for NaN gradients that could expose model internals.
    • Adversarial Inputs: Simulates noisy inputs to assess robustness.
  • Security Note: Use in controlled environments—exposing vulnerabilities could aid attackers if not mitigated.

There aren’t any published security advisories