Privacy is a system-design concern, not a final compliance checkbox. Plan for it early because user impersonation, data-management controls, explainability interfaces, and regulatory review can become expensive after implementation. section 13.3.2
At deployment, privacy depends on where predictions run and who can access the model or its outputs. Backend-contained systems have a different exposure from models running on users’ devices, where reverse engineering and misuse require stronger protection. section 15.1
Use privacy as a practical checkpoint across the ML lifecycle:
- Identify applicable regulations and involve Legal or Compliance before implementation choices become difficult to change. section 13.3.2
- Check whether training platforms protect sensitive data, especially in healthcare and legal applications. section 10.2
- Review the target platform, model exposure, APIs, and user access during serving and inference design. section 15.1 section 15.6
- Treat privacy as a tradeoff alongside latency, throughput, scalability, cost, reliability, and flexibility rather than optimizing one dimension in isolation. section 15.6
A platform decision with privacy consequences
A training platform for healthcare data may offer strong experiment tracking and inexpensive cloud execution, but privacy remains a nonfunctional requirement. A hybrid toolchain can be reasonable only when integration preserves the required controls and does not expose sensitive data unnecessarily. section 10.2
