Joint Linear and Nonlinear Computation across Functions for Efficient Privacy-Preserving Neural Network Inference

Abstract: While it is encouraging to witness the recent development in privacy-preserving Machine Learning as a Service (MLaaS), there still exists a significant performance gap for its deployment in real-world applications. We observe the state-of-the-art frameworks follow a compute-and-share principle for every function output where the summing in linear functions, which is the last of two steps for function output, involves all rotations (which is the most expensive HE operation), and the multiplexing in nonlinear functions, which is also the last of two steps for function output, introduces noticeable communication rounds. Therefore, we challenge the conventional compute-and-share logic and introduce the first joint linear and nonlinear computation across functions that features by 1) the PHE triplet for computing the nonlinear function, with which the multiplexing is eliminated; 2) the matrix encoding to calculate the linear function, with which all rotations for summing is removed; and 3) the network adaptation to reassemble the model structure, with which the joint computation module is utilized as much as possible. The boosted efficiency is verified by the numerical complexity, and the experiments demonstrate up to 13x speedup for various functions used in the state-of-the-art models and up to 5x speedup over mainstream neural networks.