CyberCertPrepAI Security FundamentalsAdversarial ML & Model AttacksEASYFree question
A model trained on ImageNet is directly applied to a medical X-ray classification task without fine-tuning. An adversary crafts an adversarial perturbation using the ImageNet model and applies it to X-ray inputs. The X-ray classifier is also fooled. What property of adversarial examples does this demonstrate?
- A.Transferability - adversarial examples crafted on one model often fool other models with different architectures and training data
- B.Universality - the perturbation works on any input regardless of the source model
- C.Certified robustness - the perturbation is within the L2 robustness radius
- D.Data poisoning - the ImageNet training set contaminated the X-ray model
Why A is correct
Transferability is the property that adversarial examples crafted against one model (the surrogate) often remain adversarial when applied to a different model (even with different architecture or training data), enabling black-box attacks via surrogate models. Universality refers to input-agnostic perturbations. Certified robustness means the model's prediction is guaranteed not to change, which is the opposite of what happened. No data poisoning is involved.
Know someone studying for AI Security Fundamentals? Send them this one.