Skip to content
AI Security

Interview question · Adversarial ML & Model Privacy

How much poisoned data does an attacker need, and why does that matter at web scale?

The intuition that a huge dataset dilutes poison turns out to be wrong for some attacks. The scored answer cites the evidence, explains why, and moves the defence to data provenance and testing.

hardNewPreparation tags: Anthropic, Google DeepMind, OpenAIUpdated Nov 2026

The intuition that a huge dataset dilutes poison turns out to be wrong for some attacks. The scored answer cites the evidence, explains why, and moves the defence to data provenance and testing.

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 training a language model on a large web-scale corpus asks how much poisoned data an attacker would need to affect their model, assuming it would take a large share.

  1. Why might less poison be needed than intuition suggests?
  2. How could an attacker get documents into a web-scale dataset?
  3. What if the backdoor in the study was simple?
  4. Which pipeline controls reduce the risk?
  5. How would you test for backdoors after training?
  6. What would you tell 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 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?Next in this trackExplain DP-SGD, its privacy and utility trade-off, and when differential privacy is the wrong tool.
DISCUSSION · 0

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