Hello To Privacy

ABSTRACT This paper introduces new protocols for secure multiparty computation (MPC) leveraging Discrete Wavelet Transforms (DWTs) for computing nonlinear functions over large domains. By employing DW Ts, the protocols significantly reduce the overhead typically associated with Lookup Table-style (LUT) evaluations in MPC. We state and prove foundational results for DWT-compressed LUIs in MPC, present protocols for 9 of the most common activation functions used in Mib, and experi- mentally ovalate the performance of our protocols for laroedo- Ryan Henry based on these results (the four protocols arise from combi- nations among two DWTs and two methods for preparing query vectors.) The new protocols are fast, have low round complexity, and give highly accurate function computations. While our new methods are highly general, we focus our experimental evaluation on a specific use case: nonlinear ac- tivation functions in deep neural networks. Prior work |11, 34] underscores the need for fast, high precision evaluation in this context, and our methods are well-suited to meet this need.