SECAI Core · phase 6 of 15
Model validation
The process of testing a trained AI model on previously unseen data to evaluate how well it generalizes (accuracy, fairness, robustness, etc.) and to confirm that it meets the requirements before deployment
The Explain card
- Plain English
- Model validation tests a trained model on data it has never seen to measure how well it generalises, checking accuracy, fairness and robustness, and confirming it meets requirements before deployment.
- Example
- Before a malware classifier goes live, the team evaluates it on a held-out set of recent samples, including families absent from training. Accuracy on the old data looked excellent; on the new set it drops sharply, and the model goes back for more work.
- Why it matters
- A model that aces its training data may fail the moment it meets reality. Validation is the evidence that it works beyond the lab, and it is where security-specific tests like adversarial inputs belong.
- Hook
- Validation is the exam on questions the model did not study for.
Where it sits in the deck
Phase 6: Model Quality, Ethics, and Responsible AI Principles
Before deploying a model, understand the quality and ethical dimensions — bias, fairness, explainability, and the human oversight needed to govern outputs responsibly.
