Privacy and compliance

How this question breaks down

1. 1ConsentCore23 minEvery row a model trains on needs a reason to exist and an expiry date. The real engineering is making a withdrawal or a deletion reach the logs, feature tables, training snapshots and models built from that row.
2. 2AnonymizationCore20 minDropping names and hashing IDs gives pseudonymous data, which is still personal data. A handful of ordinary facts, like a birthday, a postcode or four points on a map, singles most people out again.
3. 3DPCore23 minAdd calibrated noise so that any one person's data barely changes what comes out. ε measures "barely", it adds up across releases, and it is paid for in accuracy, most of all for small groups and rare examples.

Work through the parts in order and you’ve covered the full design. Each part is one topic.