TL;DR:
0. Set network to zeros
1. Probe a single input via a weight from input -> layer 1, neuron 0
2. Create a lower-than function with a bias weight and relu: relu(-input + bias) <=> 0 if weight lower than bias, >0 if bigger than bias
3. Forward layer 1, neuron 0 to output layer -> will trigger if >0
4. Repeat for all inputs and multiple biases to get an oracle for all pixels

[Full report](http://sigflag.at/blog/2022/google_ctf_2022_ocr/)