Rating:
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.
