Course / III. Development / 3.3
Building and training with privacy in mind
9 min read
Privacy-enhancing technologies (PETs)
| Technique | What it does | Trade-off |
|---|---|---|
| De-identification / pseudonymization | Removes or replaces direct identifiers | Re-identification risk remains |
| Anonymization | Irreversibly prevents identification | Hard to achieve; may cut usefulness |
| Differential privacy | Adds calibrated noise so outputs reveal little about any one person | Accuracy cost |
| Federated learning | Trains across devices without centralizing raw data | Complexity; updates can still leak |
| Synthetic data | Generates artificial data with similar statistics | May miss rare cases or leak if overfit |
| Homomorphic encryption / secure multi-party computation | Computes on encrypted or split data | Heavy compute cost |
Model choice
Pick a model type that fits the explainability need. If a decision must be explained to individuals (credit, hiring), an inherently interpretable model with acceptable accuracy may be the right choice over a black box.
Build hygiene
- Version control for data, code and models
- Experiment tracking so results can be reproduced
- Secure pipelines: access control, integrity checks on data, supply-chain review of third-party models and libraries
- Design decision log: what was chosen and why
Documentation
Model cards describe intended use, out-of-scope uses, performance across groups, limitations and ethical considerations. Together with datasheets, they are the evidence regulators and auditors ask for. If it isn't documented, assume a regulator will treat it as not done.
Check your understanding
Check 1 of 2Domain III. Development
Which technique trains a model across many phones without collecting their raw data centrally?
Check 2 of 2Domain III. Development
What does a model card describe?
Answer the checks, then mark this lesson complete
Completion and scores are saved in this browser.