Skip to content
AI Security

Interview question · Adversarial ML & Model Privacy

Is anonymising training data enough to stop privacy attacks on a model?

Removing names removes the obvious identifier and leaves the rest. The scored answer explains re-identification, why models amplify it, and what actually bounds the risk.

hardNewPreparation tags: Apple, Google, MetaUpdated Nov 2026

Removing names removes the obvious identifier and leaves the rest. The scored answer explains re-identification, why models amplify it, and what actually bounds the risk.

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 company removed names and email addresses from customer support data, trained a model on it, and now describes the model as privacy-safe.

  1. Why does removing direct identifiers not make data anonymous?
  2. What extra risk does training a model add?
  3. What if the free text contains details customers typed?
  4. How would you test the trained model?
  5. When is differential privacy worth it here?
  6. How would you make the claim defensible?

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 trackExplain DP-SGD, its privacy and utility trade-off, and when differential privacy is the wrong tool.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?
DISCUSSION · 0

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