Skip to content
AI Security

Interview question · Adversarial ML & Model Privacy

You suspect a fine-tuning dataset was poisoned. How do you investigate, prove it, and decide what to do with the model?

'Clean the dataset' is not an investigation. The answer freezes evidence, traces what changed and who could change it, tests causality by retraining, and treats the pipeline as possibly compromised too.

hardNewPreparation tags: Google DeepMind, Anthropic, OpenAIUpdated Nov 2026

'Clean the dataset' is not an investigation. The answer freezes evidence, traces what changed and who could change it, tests causality by retraining, and treats the pipeline as possibly compromised too.

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

After a routine fine-tune, a customer reports that the model gives a specific wrong answer whenever a particular product name appears in the question.

  1. What do you do first?
  2. How do you find what changed?
  3. What would you search the new data for?
  4. How do you prove the cause?
  5. What if the data looks clean?
  6. How would you run and close the investigation?

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 trackWhat is transferability of adversarial examples, and why doesn't hiding a model behind an API stop adversarial attacks?Next in this trackHow do adversarial images attack vision-language models, how is that different from text in images, and does preprocessing help?Next in this trackHow would you test whether a model update increased memorisation, and does a leak that needs thousands of queries matter?
DISCUSSION · 0

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