Privacy protects information about individuals or sensitive entities from inappropriate disclosure or use. In production ML, it becomes an operational requirement across data pipelines and deployed models, where failures can create litigation, financial loss, or information leakage. section 2.2 section 10.1
Treat privacy as a cross-system concern:
- Protect data while collecting, managing, extracting, annotating, and moving it through the ML workflow. section 2.2
- Preserve privacy when serving models, alongside versioning, resource management, explainability, and security controls. section 10.1
- Account for privacy failures as both business risks and system-design failures rather than treating them as isolated data issues. section 2.2 section 10.1
Production model serving
A deployed model must be protected against leaking sensitive information while serving predictions, making privacy part of deployment reliability rather than only data preparation. section 10.1
