Unit 11: Secure AI Deployment & Model Robustness
Scenario: A facial recognition AI at a major retailer produced wildly inaccurate results after a UI redesign on input cameras.
Root Cause:
Model drift + no post-deployment testing
Cameras fed grainier images, but the model wasn’t retrained
Mitigation:
Reintroduced regular input quality checks
Added model drift alerts
Re-trained with new data pipeline
📌 Trust was regained — after almost being lost.
