CyberCertPrepAI Security FundamentalsAdversarial ML & Model AttacksEASYFree question
A university lecturer teaching AI security wants to demonstrate availability attacks on ML services. Which attack primitive is the most direct example of deliberately degrading ML service availability?
- A.Executing a model extraction attack via repeated queries
- B.Performing a membership inference attack on the model API
- C.Constructing universal adversarial perturbations
- D.Sending sponge examples that maximize inference latency
Why D is correct
Sponge examples directly target service availability by engineering inputs that consume excessive compute per inference call. Membership inference is a privacy attack. Universal perturbations cause misclassification rather than availability degradation. Model extraction (stealing the model via queries) is a confidentiality/IP attack, not primarily an availability attack.
Know someone studying for AI Security Fundamentals? Send them this one.