Study differential privacy

1. 01

   Concepts & system design

   60 min· read · map the mechanismtimer

   R18 — Opacus documentation: PrivacyEngine, clipping, epsilon/delta

   R17 — NIST Privacy Framework

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

2. 02

   Project implementation

   120 min· build · integrate · debugtimer

   Study Differential Privacy (DP-SGD): dataset adjacency, per-sample gradient clipping, Gaussian noise multiplier, (ε, δ) privacy budget accounting, and utility degradation.

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

3. 03

   Security lab

   60 min· attack your own systemtimer

   Run a documented Opacus (or equivalent) DP training experiment on a small local dataset; compare accuracy and loss across two ε budgets vs. non-private training. *(Do not claim production privacy from a tutorial).*

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

4. 04

   Verification & evidence

   30 min· test · record evidencetimer

   DP experiment report comparing (ε, δ) settings vs. task utility.

   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

• 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.

• docsOpacus (official docs)
  Tutorials

  Official DP-SGD tutorials wiring PrivacyEngine, per-sample gradient clipping, noise multiplier, and epsilon accounting into a real training loop.

  R18checked 2026-11-11
• videoPyTorch Foundation
  Differential Privacy on PyTorch | PyTorch Developer Day 2020

  Explains the PrivacyEngine attach step and what clipping plus Gaussian noise do to the gradient, including the utility cost of a tighter epsilon.

  R18checked 2026-11-11
• articleOpenMined
  Differentially Private Deep Learning in 20 lines of code

  A minimal end-to-end DP-SGD run that prints the privacy budget spent, useful as the smallest reproducible harness for comparing epsilon against accuracy.

  R18checked 2026-11-11
• repoMeta PyTorch (GitHub)
  meta-pytorch/opacus

  Source and examples for the accountant and per-sample gradient machinery, including how the noise multiplier is derived from target epsilon and delta.

  R18checked 2026-11-11
• articleMukul Rathi
  Deep Learning with Differential Privacy (DP-SGD Explained)

  Walks dataset adjacency, L2 clipping of per-sample gradients, and the noise-multiplier-to-epsilon relationship with the math, which is what the day asks you to reason about rather than just call.

  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.