A straight-to-the-point implementation of homomorphic encryption on neural networks using Efficient Integer Vector Homomorphic Encryption. This allows a client to use a neural network without revealing the input and output to the network. A good introduction on building encrypted neural networks can be found at https://iamtrask.github.io/2017/03/17/safe-ai/
Currently there is no working non-linearity, so the neural network is a linear perceptron.