Skip to content
AI Security

Interview question · Adversarial ML & Model Privacy

What is clean-label poisoning, and why is it harder to catch than label flipping?

Every poisoned example looks correctly labelled to a reviewer. The scored answer explains the mechanism, why label audits miss it, and the defences that work at the pipeline level.

hardNewPreparation tags: Google DeepMind, Meta, AppleUpdated Nov 2026

Every poisoned example looks correctly labelled to a reviewer. The scored answer explains the mechanism, why label audits miss it, and the defences that work at the pipeline level.

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 wildlife image classifier is trained partly on images contributed by volunteers, which reviewers check for correct labels before they enter the training set.

  1. How can poisoning work if every label is correct?
  2. Why does the label review miss it?
  3. Which detection methods can help?
  4. What if contributors can submit unlimited images?
  5. How would you test the trained model?
  6. What would you put in place for contributed training data?

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 trackHow much poisoned data does an attacker need, and why does that matter at web scale?Next in this trackModel inversion, model extraction and attribute inference: what is the difference?Next in this trackIs anonymising training data enough to stop privacy attacks on a model?
DISCUSSION · 0

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