Track 07 · Data & governanceLesson 05 of 11
Finding and masking personal data
Discover, mask and tokenise PII before it stalls the project.
- Specialisation · Healthcare
- 30 min read
- 75 min hands-on
- Pro
What you'll be able to do
01- Explain how PII, GDPR special-category data, HIPAA PHI and PCI cardholder data differ, and which de-identification standard applies to each.
- Choose between redaction, masking, pseudonymisation, vault tokenisation, format-preserving encryption and anonymisation for a field and its consumer.
- Implement detectors that combine patterns, checksums and context, and measure precision and recall per entity type.
- Design where masking happens in an LLM pipeline and where un-masking is allowed.
- Defend the residual re-identification risk to a CISO and a privacy officer with numbers.
Before you start
- Finding where the data really lives and Is the data ready?: you need the data map and the platform before you can attach policies.
- Guardrails the risk team can sign: PII detection as a runtime guardrail.
- Retrieval that respects who can see what: ACLs decide who sees a document; this concept decides which values inside it anyone sees.
The short version
Enterprise AI stalls on PII because nobody can show which sensitive values cross which boundary. The fix is data engineering, not a prompt: discover sensitive data in columns and free text, classify it by regime, measure your detectors, and transform values near the source. Deterministic keyed tokens keep joins and reasoning working; a vault the model never touches holds the mapping; un-masking happens only at the edge, for an authorised viewer, logged.
- Classification drives treatment. A card number, an NHS number and a diagnosis fall under different rules.
- Detectors are classifiers. Report precision and recall per entity type; an overall score hides the entity you never detect.
- Pseudonymised is not anonymous. Under GDPR Recital 26, tokenised data is personal data for whoever can reverse it.
- Mask early, un-mask late. Before embedding, prompting, logging and tracing; reveal only in the render layer or a trusted tool.
- The vault is a security system. Role, purpose and case scope gate every reveal, and the log is audit evidence.
Pro lesson
The rest of this lesson is in Pro.
The mental model, worked example, field note, failure table, exercise and self-check, with the voice tutor.
- Every stage and specialisation, in full
- Every lab and every role project
- Halden weeks 2–10, on your record
- Every drill, live in the playground
- All solutions, plus the server judge
- Lab Assistant and every Teach with AI