Jin, C., Aung, KMM., Xin, Z. Secure Collaborative Design of Experiments with Homomorphic Encryption. HomomorphicEncryption.org Workshop
Abstract:
Given a neural network model that computes the relationship f between an input X, an output Y = f(X) and a desired output Y*, the Design-of-Experiments (DoE) problem asks for some X’ with f(X’) close to Y*. Consider a model owner who offers to run a DoE algorithm whose result improves with more iterations, and the user who wants an input value for a desired Y* but is unwilling to reveal Y*. We propose a private DoE protocol that addresses the dilemma using fully homomorphic encryption. The user encrypts Y* and an initial input value X0 into [Y*] and [X0] and sends the ciphertexts to the model owner; the model owner converts the DoE algorithm to an equivalent function operating on the encrypted data and returns a result [X’] and the loss [l] in ciphertexts which the user can decrypt. The interaction continues for multiple rounds if the user hopes to further reduce the loss. As a proof of concept, we tested the protocol on neural network models for two simulated datasets, with the DoE algorithm being the backpropagation algorithm optimizing the input values by gradient descent. The protocol achieved a small mean squared error loss within a few iterations.
License type:
Publisher Copyright
Funding Info:
This research / project is supported by the A*STAR - RIE2020 Advanced Manufacturing and Engineering (AME) Programmatic Programme
Grant Reference no. : A19E3b0099