A bank deploys an AI model to approve credit card transactions. The model is hosted on-premises and only the final decision (approve/decline) is exposed via an API. An attacker wants to craft a transaction that is approved but would normally be declined. Which attack vector is most feasible given the attacker's access?

White-box gradient attack using the model's training data.

Data poisoning attack on the training pipeline.

Model inversion attack to reconstruct the training data.

Black-box evasion attack using the API's decision as an oracle to estimate gradients.

A company is deploying an AI model for medical diagnosis. They must comply with data privacy regulations and also protect against adversarial attacks. Which combination of strategies best addresses both concerns?

Adversarial training and differential privacy

Gradient masking and data poisoning

Model pruning and membership inference

Input sanitization and model inversion

A security team is testing a facial recognition system that grants access to a data center. They want to prove that an attacker could bypass the system using a printed photo. Which approach best demonstrates this risk while staying within a realistic threat model?

Generate a universal adversarial perturbation and apply it to a photo of an authorized employee, then print it.

Apply a random noise filter to the photo and hope the model misclassifies it.

Use a gradient-free attack that requires thousands of queries to the live system, then print the result.

Enlarge the photo and print it in high resolution to trick the model.

Which of the following is an example of an attack vector that threat modeling might identify for an AI system?

Increasing the number of layers in the neural network.

Adding a new feature to the model's training data.

Crafting inputs with imperceptible perturbations to cause misclassification.

Optimizing the model's hyperparameters for faster training.

EC-CouncilCertified Offensive AI Security Professional

Domain 4Objective 1

Adversarial Machine Learning Techniques COASP Practice Questions (Page 2)

Part of the Adversarial Machine Learning and Model Privacy Attacks domain, which makes up ~13% of our current practice bank.

38questions here

8free pages

3concepts

Questions 6–10

6application · medium
A bank deploys an AI model to approve credit card transactions. The model is hosted on-premises and only the final decision (approve/decline) is exposed via an API. An attacker wants to craft a transaction that is approved but would normally be declined. Which attack vector is most feasible given the attacker's access?
Select an answer first
7expert · hard
A company is deploying an AI model for medical diagnosis. They must comply with data privacy regulations and also protect against adversarial attacks. Which combination of strategies best addresses both concerns?
Select an answer first
8application · medium
A security team is testing a facial recognition system that grants access to a data center. They want to prove that an attacker could bypass the system using a printed photo. Which approach best demonstrates this risk while staying within a realistic threat model?
Select an answer first
9foundation · easy
Which of the following is an example of an attack vector that threat modeling might identify for an AI system?
Select an answer first
10application · medium
A company deploys a chatbot that uses a large language model. They are threat-modeling and identify that an attacker could craft prompts to make the chatbot reveal sensitive information. Which attack vector is this?
Select an answer first

Finished these 5 questions?

Review the revealed explanations, or continue through the curriculum.

Next 5 questions →Back to COASP

Free Basic Practice is a study aid with revealable answers — not a scored exam. Examers.io is independent and not affiliated with or endorsed by EC-Council. “COASP” is a trademark of its owner, used for identification only.

Adversarial Machine Learning Techniques COASP Practice Questions (Page 2)

Questions 6–10

A company is deploying an AI model for medical diagnosis. They must comply with data privacy regulations and also protect against adversarial attacks. Which combination of strategies best addresses both concerns?

A security team is testing a facial recognition system that grants access to a data center. They want to prove that an attacker could bypass the system using a printed photo. Which approach best demonstrates this risk while staying within a realistic threat model?

Which of the following is an example of an attack vector that threat modeling might identify for an AI system?

A company deploys a chatbot that uses a large language model. They are threat-modeling and identify that an attacker could craft prompts to make the chatbot reveal sensitive information. Which attack vector is this?

Keep studying — Adversarial Machine Learning and Model Privacy Attacks