I recently read about "functional encryption" which seems interesting, although I didn't understand yet how it works ... but is it possible to combine it or adapt it with homomorphic encryption in the context of neural networks particularly.
Suppose you have a neural network that you run over some data encrypted homomorphically (say paillier for example) .. linear computations will be performed without any problem .. but the only issue will be with activation functions which are non linear, and thus require to decrypt the intermediate results (weighted sums) in order to apply the activation functions ...
in case we don't want to disclose the private key, can we use something like functional encryption in order to allow the application of activation functions only, something like a special private key that allows only a specific activation function to be performed, and do not allow to decrypt any other information ?!