[Original writeup](https://imp.ress.me/blog/2022-07-05/google-ctf-2022-ocr/)

Given a toy neural network being run against images of flag text, craft model weights to turn ReLU layers into a comparison oracle to leak image brightness data and reconstruct image dataset locally. More efficient solutions exist involving step-wise activation of output layers and crafting convex hulls.

Original writeup (https://imp.ress.me/blog/2022-07-05/google-ctf-2022-ocr/).