Practice tests · 40 items

Adversarial ML & Model Privacy practice test

The attacker model for ML, evasion and adversarial examples, availability, targeted and backdoor poisoning, poisoning at web scale, model extraction, membership inference, training data extraction, model inversion, differential privacy and privacy auditing. The pool for this topic holds 10 easy, 18 medium and 12 hard items. Each answer comes with its reasoning and a link to the question or concept page it was drawn from.

Configure a test

Topic

Difficulty

Questions

10 in this pool · about 7 min

Reveal answers

Sample items with answers

easy · sample

In machine learning, what is an adversarial example?

A test case written to show a model works

✓An input changed slightly, often invisibly, so that a model gets it wrong

A training example that has been deliberately mislabelled so the model learns the wrong answer for it

An input written in another language to confuse the model's tokenizer

Szegedy and colleagues showed in 2013 that tiny, targeted perturbations can flip a classifier's output while a human sees no change. Attacks such as FGSM and PGD compute them from gradients. The model was never trained to be stable in those directions.

easy · sample

What separates an evasion attack from a poisoning attack?

✓Evasion changes inputs at inference; poisoning changes the training data

Nothing important

Evasion targets open-weight models only, while poisoning only works against models behind an API

Evasion steals the model, while poisoning deletes it from the server

Evasion attacks craft inputs to a trained model, such as a modified image or a reworded spam email. Poisoning attacks act earlier, inserting or altering training examples so the model learns the attacker's behaviour. They need different controls: input robustness versus data provenance.

After the test

Choosing from options is easier than reasoning it out in front of a panel. These pages make you reason it out: