Privacy and ML: Index
| From Machine Learning | |
| Privacy and ML | |
|---|---|
| Page metadata | |
| First created | May 31, 2026 |
| Last edited | May 31, 2026 |
Prep plan — 24 hours
HM call Monday 3
PM EDT, 30 minutes, WebEx, with Erik. First-round fit conversation: team, role, background. Not a technical screen. The intake survey already put DP-SGD, FedAvg, secure aggregation, ε/δ-DP, local vs central vs shuffle DP, Opacus, Flower, Core ML, vLLM, FAISS, and sentence-transformers on the record by name — bar is conversational depth on each, not implementation depth.What gets evaluated
- Can I talk about my own work clearly. Most of the call.
- Do I understand the problem space (privacy-preserving ML) at PM/architect level — vocabulary, tradeoffs, where each tool fits.
- Do I seem like someone he wants on the team — communication, energy, curiosity.
- Do my goals match the rotation (the intake survey ranked Growth & Learning #1; expect a probe on whether that’s real).
- My questions for him. Last 5–10 minutes.
What does NOT get evaluated
Math derivations. Code. Whiteboarding DP-SGD. That’s later rounds, if at all.
The four prep areas
1. DP / FL vocabulary tour at conversational depth.
60 seconds out loud on each named term. Captured as wiki pages under Differential Privacy/ and a new Federated Learning/ folder. Terms: ε/δ-DP, local vs central vs shuffle DP, DP-SGD, FedAvg, secure aggregation, the utility–privacy–bandwidth tradeoff. Plus the tooling layer: Opacus, Flower, Core ML, vLLM, FAISS, sentence-transformers — one paragraph each on what it is, when you reach for it.
2. Apple PPML’s public story. What the team has shipped, what their architectural bets look like. Local DP in iOS keyboard (2016 launch, Erlingsson / Korolova / Apple Privacy team), Private Cloud Compute (2024), Apple Intelligence privacy architecture, on-device LLM inference. Enough to connect intelligently when Erik describes the team.
3. The Mozilla story, privacy-leading. Same content as the Atomic walkthrough, reframed: lead with the privacy property (on-device, no telemetry leaves the device) rather than the ML novelty. The 500 KB ONNX vs. 1.1 GB lookup table beat still lands; the framing is “we kept the privacy posture by going on-device, and ML is how we made on-device feasible at this accuracy.” Rehearsed aloud, target 4–5 min.
4. Three questions for Erik. About the team, the rotation, what success looks like at 4 / 12 weeks.
Schedule
Sunday (today)
- Morning / early afternoon: vocabulary tour (areas 1 and parts of 2). Socratic, captured to the wiki pages.
- Late afternoon: Apple PPML public story page.
- Evening: rehearse Mozilla story aloud, privacy-leading frame, time it. Draft the three questions for Erik.
Monday morning
- Re-read the wiki notes once. No new material. No drilling.
- Walk, eat, hydrate. Camera at eye level, door closed, water on desk.
- Sit, breathe, click join.
State on the call
Same frame as the Anson call: this is the format I have never lost in. Warm, conversational, anchored on the work I actually did, confident vocabulary on the surrounding terms. Do not over-rehearse into rigidity. Lead with concrete numbers, not the structure around them.
Pages
Index
- Differential Privacy. Reference page on differential privacy — what it is, the ε/δ definition, the three trust models (local / central / shuffle), and the tools used to apply it to machine learning.
