CyberCertPrepAI Security FundamentalsAdversarial ML & Model AttacksEASYFree question
A computer vision model is deployed for autonomous vehicle stop sign recognition. Which adversarial attack scenario represents the HIGHEST real-world safety risk?
- A.B. A targeted physical adversarial attack that consistently causes the model to misclassify a stop sign as a speed limit sign, preventing braking
- B.A. An untargeted attack causing the model to be uncertain about any classification
- C.C. A model extraction attack where a competitor copies the architecture
- D.D. A membership inference attack determining which stop signs were in the training data
Why A is correct
For safety-critical autonomous systems, a targeted attack causing a specific high-confidence wrong classification (stop sign → speed limit) has direct life-safety implications - the vehicle does not brake. This is the exact scenario motivating real-world autonomous driving adversarial research (Eykholt et al., 2018 demonstrated stop sign patches causing misclassification). Untargeted uncertainty is dangerous but less predictable in its safety outcome. Model extraction and membership inference are IP/privacy concerns with no direct physical safety impact.
Know someone studying for AI Security Fundamentals? Send them this one.