Revisit advanced privacy and deletion

1. 01

   Concepts & system design

   60 min· read · map the mechanismtimer

   R16 — Presidio

   R17 — NIST Privacy

   R18 — Opacus DP

   Read the assigned sections; map the mechanism, threat model and trade-offs.

2. 02

   Project implementation

   120 min· build · integrate · debugtimer

   Revisit advanced privacy and deletion propagation. Audit secondary caches, error traces, dead-letter queues, and embedding payloads for subtle cross-tenant or post-deletion leakage paths.

   Build, integrate, debug and improve the increment in the active project.

3. 03

   Security lab

   60 min· attack your own systemtimer

   Write a cross-store leakage test (e.g., querying cached answers or DLQ payloads after a user deletion request) and ensure D4 and D10 purge all derived artifacts.

   Test the abuse case, failure path, authorization bypass or resource limit.

4. 04

   Verification & evidence

   30 min· test · record evidencetimer

   Updated privacy threat model and passing cross-store deletion test.

   Add or run regression tests; record the 9-point evidence record.

5. 05

   Deliberate practice & Socratic review

   30 min· explain aloud · update gapstimer

   Answer the 8 Socratic questions without notes: What? Why? How? What if? Why not? How can I prove it? Where can it break? What have I learned?

   Re-explain the hardest concept aloud, inspect one edge case, update the gap log.

6. total5 h

• R16Microsoft Presidio↗ microsoft.github.io

  Analyzer, anonymizer, custom PII recognizers, and probabilistic detection false-positive/negative limits.

• R17NIST Privacy Framework↗ nist.gov

  Identify-P, Govern-P, Control-P, Communicate-P, and Protect-P lifecycle controls.

• R18Opacus Documentation↗ opacus.ai

  Differential privacy PrivacyEngine, gradient clipping, noise multiplier, $(\epsilon, \delta)$ accounting, and utility trade-offs.

Read only the named sections needed for today's task. Reading sits inside the 60-minute concepts block; it is not extra time.

• docsMicrosoft · Analyzer/anonymizer docs
  Microsoft Presidio

  Documents the analyzer recognizer pipeline and the anonymizer operators, plus how to add custom recognizers — the layer where false positives and false negatives become a documented limit rather than a surprise.

  R16checked 2026-11-11
• videoM&M Tech · 14 min
  Data Anonymization With Microsoft Presidio - Full Step By Step Tutorial

  Walks entity detection, adding a custom entity, anonymizing, and custom anonymization in sequence, which is the exact flow for extending the D4 gateway to a new PII type.

  R16checked 2026-11-11
• videoOpenMined · Conference tutorial
  OM PriCon 2020 Tutorial: Differentially Private Model Training with Opacus - Davide Testuggine

  Links DP-SGD's three mechanisms — per-sample gradient clipping, noise addition, and privacy accounting — to the PrivacyEngine API, which is what makes an (epsilon, delta) claim auditable rather than nominal.

  R18checked 2026-11-11
• articleNicchin · 14 min read
  GDPR Right to Erasure in Vector Databases: What Delete Misses

  Shows that the erasure boundary is the embedding, not the document, and that soft-deleted HNSW vectors can stay reconstructible — the secondary-cache and post-deletion leakage path you are auditing.

  R17checked 2026-11-11
• courseOpacus (Meta) · Runnable notebook
  Building an Image Classifier with Differential Privacy

  A hands-on DP-SGD run that sets max grad norm, noise multiplier, and delta, then shows ModuleValidator rejecting BatchNorm because per-sample gradients break — the utility-versus-privacy trade-off in practice.

  R18checked 2026-11-11

A lab is complete only when it produces evidence. Installing PyRIT, Garak, Presidio or TransformerLens is not completion. Saved on this device only.