Would it be possible to outsource computation to a secret node’s GPU (and therefore untrusted) for intermediate results?
I was wondering if something like this would be possible within the Enigma network: outsourcing matrix multiplication in ANN training (forward passes) from SGX to GPU:
Slalom: Fast, Verifiable and Private Execution of Neural Networks in Trusted Hardware
It’s a very interesting proposition, but we aren’t looking into mixed trusted-untrusted execution right now.
I haven’t read the paper yet but I assume the same idea could work for speeding up MPC/FHE computations as well.