312-41 practice question 13 of 35

Julianne Moore, Lead AI Systems Architect, is conducting an investigation on a facial recognition access system that recently failed a security…

Full set: 100 questions
Julianne Moore, Lead AI Systems Architect, is conducting an investigation on a facial recognition access system that recently failed a security audit. The audit team demonstrated that by wearing a specifically crafted pair of noisy pattern eyeglasses, an unauthorized user could consistently trick the system into identifying them as the CEO. Julianne confirms that the system’s source code is intact and the original database of face images used to train the model was verified as clean and unaltered. Julianne must categorize this vulnerability in her report to the CISO. Which AI-specific security threat characterizes the method used to bypass the system’s identification controls?
Answer options
A: Prompt Injection
B: Data Poisoning
C: Adversarial Attacks Correct answer
D: Model Theft
Correct Answer:
C. Adversarial Attacks
Explanation

The eyeglasses embed carefully crafted perturbations that cause the classifier to mis-identify the intruder as the CEO while leaving all other inputs unchanged. Such deliberately engineered, input-based manipulations are textbook adversarial attacks against vision models, not data-poisoning or prompt manipulation.

Why Incorrect

A. Prompt injection targets language-model prompts, not vision classifiers.
B. Data poisoning corrupts the training set; the set was verified clean.
D. Model theft extracts parameters or functionality; no theft occurred.

References

1. M. Sharif et al., “Accessorize to a Crime: Real and Stealthy Attacks on State-of-the-Art Face Recognition,” CCS 2016, pp. 1528-1540, doi:10.1145/2976749.2978392.

2. NISTIR 8269, “A Taxonomy and Terminology of Adversarial Machine Learning,” §3.2 “Evasion Attacks,” 2020.

3. Stanford CS 331B Lecture 7, “Adversarial Examples in Vision,” slides 9-12, 2022.

Scroll to Top