CyberCertPrepAI Security FundamentalsAdversarial ML & Model AttacksEASYFree question
A university lecturer explains the difference between targeted and untargeted adversarial attacks to students. Which statement correctly distinguishes the two?
- A.A targeted attack uses white-box model access; an untargeted attack uses only black-box access
- B.A targeted attack crafts a perturbation that causes the model to output a specific chosen class; an untargeted attack only requires the model to output any incorrect class
- C.A targeted attack modifies the training data; an untargeted attack modifies inference inputs
- D.A targeted attack applies to regression; an untargeted attack applies to classification
Why B is correct
In a targeted adversarial attack, the adversary crafts a perturbation so the model outputs a specific target class (e.g., misclassifying a stop sign as a speed-limit sign). In an untargeted attack, any misclassification is sufficient. Access level (white/black box) is independent of targeted/untargeted distinction. Training vs. inference time distinguishes poisoning from evasion attacks. Both types apply to classification tasks.
Know someone studying for AI Security Fundamentals? Send them this one.