Case 79BA0F · AI security · L3 Applied
Adversarial Examples (Evasion)
- 01 What is it?
- 02 How is it abused?
- 03 How do you stop it?
What a strong answer covers
Try it out loud first. Then check yourself:
- Inputs perturbed, often imperceptibly, so a model misclassifies at inference; crafted with gradients (FGSM, PGD) or queries.
- Stakes: evading malware and phishing classifiers, stickers that fool vision models, spoofed face recognition.
- Black-box works too: perturbations often transfer between models, and score-based queries can estimate gradients.
- Defences: adversarial training, input transforms, ensembles; many published defences fell to adaptive attacks.
- Never make one model the only control: layer detections, limit score exposure, and watch for probing query patterns.
If the interviewer pushes back
- Why does returning confidence scores from a malware classifier API make evasion easier, and what would you return instead?
- How do evasion attacks on a classifier relate to jailbreaks on an LLM?