pwnies-please: Adversarially Attacking an Image Classifier
For UIUCTF 2021, I led a small group of colleagues to design and
deploy a CTF challenge wherein competitors had to fool an image classifier
based on resnet18
.
The challenge asks you to upload an image to the “pwny bouncer”, who will then classify “pwny” or “not a pwny” using a naive classifier model, and a robust classifier model. To win the flag, you must submit 10 images that “fool” the nonrobust classifier (i.e. cause a mismatch between the classifiers).
See a full writeup by the prize-winning first blood solver here: