Questions: Adversarial Examples and Robustness

5 questions to test your understanding

Score: 0 / 5
Question 1 Multiple Choice

A neural network achieves 98% accuracy on a held-out test set. A researcher then applies FGSM to 100 of those correctly classified images and finds the network misclassifies 85 of them. What does this demonstrate?

AThe test set was too small to give a reliable accuracy estimate
BHigh test accuracy does not guarantee robustness — the model is brittle against adversarial perturbations
CFGSM produces unrealistic inputs that no real attacker would generate
DThe network needs more training epochs to generalize properly
Question 2 Multiple Choice

Why can adding a tiny perturbation of magnitude ε to each dimension of a high-dimensional input reliably fool a neural network, even when no single perturbed pixel is noticeable?

ABecause the perturbation shifts the input into a different data distribution that the model has never seen
BBecause the perturbations accumulate: the total effect on the output can be as large as ε × d, where d is the input dimensionality
CBecause neural networks only process a small subset of input dimensions at a time
DBecause ε-perturbations happen to target the most important pixels as identified by the gradient
Question 3 True / False

Adversarially trained models typically achieve lower accuracy on clean, unperturbed test images than models trained without adversarial examples.

TTrue
FFalse
Question 4 True / False

An adversarial perturbation is expected to be visible to the human eye in order to reliably fool a state-of-the-art neural network classifier.

TTrue
FFalse
Question 5 Short Answer

Explain why high-dimensional input spaces make neural networks particularly vulnerable to adversarial perturbations, even when those perturbations are small in any single dimension.

Think about your answer, then reveal below.