1. 9
  1. 1

    An interesting approach, but what I miss from this blog post is how the learning would actually be done. Can you calculate the gradients from the encrypted output? Is it possible that the network can perform well with weights that have been trained on the unecrypted input before?

    Additionally, the error of the operations is not reported. From what I recall, performing homomorphic operations always leads to noise in the output, which will increase considerably via error propagation. I would imagine that this would be significant in a neural network, although the size of this one does not seem too big, so perhaps it would be okay.

    Lastly, and this is a mistake on my side, I thought ML would be used for homomorphic encryption and not that homomorphic operations would be used on the operations inside of the net. But that was just due to me skimming the title.