FGSM and PGD Attacks on Image Classifiers COASP Practice Questions (Page 2)

Questions 6–10

1. 6foundation · easy

   When visualizing an adversarial example alongside the original image, what is the most common way to highlight the perturbation?

   • ADisplay the model's confidence scores for both images
   • BDisplay the gradient of the loss with respect to the model weights
   • CDisplay the absolute difference between the adversarial and original images
   • DDisplay the original image with a color overlay of the predicted class
   Reveal Solution Discussion Select an answer first
2. 7application · medium

   A researcher is running a PGD attack and notices that the adversarial image after the first iteration is already outside the epsilon-ball. What should they do to fix this?

   • AIncrease the epsilon value
   • BReduce the step size to zero
   • CUse FGSM instead
   • DApply the projection step after each iteration to clip the perturbation back into the epsilon-ball
   Reveal Solution Discussion Select an answer first
4. 8application · medium

   A data scientist is preparing a report on adversarial robustness. They generate adversarial examples using FGSM and want to visually demonstrate the perturbation. Which visualization approach best shows the difference between the original and adversarial images?

   • APlot the model's loss curve during training
   • BDisplay the original and adversarial images side by side and compute the pixel-wise absolute difference, scaling it for visibility
   • CDisplay only the adversarial image without the original
   • DShow the model's confidence scores as a bar chart
   Reveal Solution Discussion Select an answer first
5. 9application · easy

   A developer is implementing FGSM in PyTorch. After computing the loss and calling loss.backward(), which tensor holds the gradient needed to craft the adversarial example?

   • Amodel.weight.grad
   • Binput_image.grad
   • Closs.grad
   • Doptimizer.grad
   Reveal Solution Discussion Select an answer first
6. 10application · easy

   A security team is evaluating the robustness of a production image classifier. They run FGSM with epsilon=0.1 on a 10,000-image test set and find that 8,500 images are misclassified. They want to report the attack success rate. What is the correct success rate?

   • A15%
   • B85%
   • C8.5%
   • D0.85%
   Reveal Solution Discussion Select an answer first