Skip to content
AI Security

Interview question · Adversarial ML & Model Privacy

Explain DP-SGD, its privacy and utility trade-off, and when differential privacy is the wrong tool.

DP gives a real guarantee with a real cost. The scored answer explains clipping, noise and accounting, then compares DP with cheaper controls for a concrete dataset.

expertNewPreparation tags: Apple, Google DeepMind, MetaUpdated Nov 2026

DP gives a real guarantee with a real cost. The scored answer explains clipping, noise and accounting, then compares DP with cheaper controls for a concrete dataset.

Make your first attempt

What would you say before reading the answer?

  1. 01 State the assumptions your answer depends on.
  2. 02 Explain the mechanism and choose a control.
  3. 03 Name a condition that would change your decision.

The worked answer includes an approach, scoring criteria and follow-up questions. Compare the reasoning with your own.

10 answers per topic instead of 5, plus saved progress and bookmarks · costs nothingor unlock all 529 remaining answers · $25

Every worked answer in all 17 topics, every lesson in every course, the full concept curriculum, the complete practice-test bank, and the downloadable security field guides.

6 months · One payment · No auto-renewal

What if · the follow-ups in this answer

The answer works through 6 follow-ups like an interviewer would

A team plans to train a diagnostic model on hospital records with differentially private SGD and asks what it will cost and whether it is the right tool.

  1. How does DP-SGD work?
  2. What does it cost?
  3. What if one patient has many records?
  4. What if the concern is population-level inference?
  5. When would cheaper controls be enough?
  6. What would you advise the team?

Each rung has a nudge, a model answer, the principle that carries over and the weak answer to avoid. Premium opens every answer and ladder.

UP NEXT ON YOUR JOURNEY
Next in this trackA deployed model may be leaking its training data. How do you respond?Next in this trackWhat can watermarking of model outputs or weights prove, and what can it not?Next in this trackWhy is membership inference hard even to define for LLMs trained on internet-scale data?
DISCUSSION · 0

No comments yet. Share a correction or a different way to answer it.