Skip to content

Topic Drill 6.1.4: Adversarial Testing of ML Systems ​

About this drill

LO AI-6.1.4 (K2): Summarize adversarial testing of machine learning systems.
Lesson: Ch 6, section 6.1.4

This drill covers every point in section 6.1.4:

  • Perturbations (often imperceptible), adversarial examples, slightly modified legitimate inputs
  • Purpose: safeguards and robustness; accidental examples vs. attacks; complexity and evolving techniques
  • Black-box (equivalent model + transferability, or brute force), white-box, manual vs. automated

Topic Drill: 6.1.4

Q1. Researchers place a few small stickers on a stop sign. Humans still clearly see a stop sign, but the car's classifier confidently reports "speed limit 80". What is the modified sign, in the syllabus's terms?

AA data drift event
BA metamorphic relation
CAn adversarial example
DAn outlier in the training data

Q2. What is the main benefit of identifying vulnerabilities through adversarial testing?

ADevelopers can incorporate safeguards and make the model more robust against adversarial examples, whether met accidentally or used in attacks
BIt proves that the model has no other defects
CIt replaces the need for a test dataset
DIt increases the model's accuracy on the training data

Q3. How does the syllabus distinguish an "adversarial attack" from adversarial examples the model might simply encounter?

AAn attack uses larger perturbations
BAn attack only targets GenAI systems
CAn attack happens only during training
DAn adversarial attack involves the malicious use of adversarial examples

Q4. Testers have no access to a deployed model's internals. They train their own similar model, craft adversarial examples using its known internals, and then use those inputs against the deployed model. What assumption does this rely on?

AThat the deployed model is overfitted
BTransferability: equivalent models share classification boundaries
CThat the deployed model is locked
DThat both models use the same training data

Q5. Without building any equivalent model, a team sends millions of random variations of inputs to a black-box model, hoping some will be misclassified. Which approach is this?

AWhite-box adversarial testing
BMetamorphic testing
CA brute force approach to black-box adversarial testing
DBack-to-back testing

Q6. Why does WHITE-box adversarial testing typically make it easier to craft adversarial examples?

ABecause understanding the model's internals (architecture, parameters, training process) makes it easier
BBecause white-box tests need no inputs
CBecause white-box models are always smaller
DBecause white-box testing uses only manual methods

Q7. Which ongoing challenge of adversarial testing does the syllabus mention?

AAdversarial examples are always visible to humans
BIt can only be used on image classifiers
CIt requires labeled data for every possible input
DGenerating effective test inputs is technically complex, and keeping up with evolving attack techniques is an ongoing challenge

Q8. Which TWO statements about adversarial testing are CORRECT?

Select 2 options.

AIt can only be performed as black-box testing
BIt can be done manually by crafting specific examples, or with automated algorithms that generate many variations
CIt re-runs previously passed tests unchanged to avoid failures
DIts test inputs are often slightly modified versions of legitimate inputs
EIt applies only to generative AI

0 / 8 answered

← 6.1.3 Statistical Performance Testing · Next drill: 6.1.5 Metamorphic Testing →

Unofficial study guide for the ISTQB® Certified Tester AI Testing (CT-AI) exam. Not affiliated with International Software Testing Qualifications Board (ISTQB®).