SECAI Core · phase 15 of 15
Data drift
A phenomenon where the statistical properties of input data change over time, potentially reducing the accuracy and reliability of machine learning models or analytical systems.
The Explain card
- Plain English
- Data drift happens when the statistical properties of the input data a model receives change over time, so the model is seeing a world it was not trained on and its accuracy quietly declines.
- Example
- A phishing detection model trained on human-written email is increasingly fed messages written by generative AI. The vocabulary and structure of incoming mail shift, and detection rates fall without any change to the model.
- Why it matters
- Drift degrades defences silently and can be induced deliberately by attackers who feed a system unusual inputs. Monitoring input distributions is how you notice.
- Hook
- The map is the same, but the territory moved.
Where it sits in the deck
Phase 15: MLOps, Continuous Delivery, and Operational Resilience
Close the loop — operational practices, drift management, and delivery discipline that keep deployed AI systems healthy, current, and continuously improving.
