A Communication-Reduced Privacy-Preserving ViT Inference Framework for Distributed Edge Intelligence

Abstract

To address privacy leakage from user images, intermediate representations, and outputs during Vision Transformer (ViT) inference in distributed edge services, this paper presents SViT, a two-server secret-sharing framework with offline correlated randomness. The revised design specifies fixed-point arithmetic over the ring Z_(2^64) with 16 fractional bits, fresh one-time masks, probabilistic truncation, numerical ranges, and complete input-output procedures for SExp, SDiv, SSqrt, SVar, SLayerNorm, SSoftmax, and SGeLU. The protocols use fixed-depth range reduction, lookup-assisted initialization, and a constant number of Newton updates, so their online depth is independent of numerical convergence tolerances. Under the semi-honest, non-colluding-server model, the revised security analysis defines approximate ideal functionalities, public leakage, simulator inputs, and sequential composition. Existing microbenchmarks show 2.28-6.50 times lower runtime and 4.00-14.20 times lower online communication than CrypTen for the reported core operators; for SDiv, runtime decreases from 9.1 ms to 1.4 ms and communication from 10.8337 MB to 0.7629 MB. Synthetic Q16 numerical checks report maximum absolute errors of 5.10×10−5 for SExp and 4.73×10−4 for SGeLU, maximum relative errors of 1.27×10−5 for reciprocal and 3.72×10−5 for square root, and 100% top-1 consistency over 10,000 randomly generated Softmax vectors. The reported end-to-end result remains 3799.744 ms and 2.85 GB per inference; therefore, SViT is described as communication-reduced relative to the evaluated MPC baselines rather than universally lightweight, and its present practical scope is primarily high-bandwidth LAN or provider-edge deployments.