Paper 2025/1315

NetAdapt: Network-Adaptive Hybrid Protocol Assignment for PPML

Abstract

The widespread use of machine learning on sensitive data makes Privacy-Preserving Machine Learning (PPML) essential for data confidentiality. Current state-of-the-art PPML systems employ hybrid protocol designs to evaluate various operators to achieve better performance; yet, existing hybrid approaches adopt fixed protocol assignments without considering the deployment setting, resulting in inefficiencies across diverse network environments, such as LANs and WANs. To address this, we introduce NetAdapt, a cost-model-driven framework that automatically assigns FHE and MPC protocols to optimize inference efficiency under varying network conditions, by designing a predictive cost model based on the dialect of the Multi-Level Intermediate Representation (MLIR)’s Tensor Operator Set Architecture (TOSA), and an ILP-based solver for managing the cross-protocol conversion while dynamically updating FHE costs based on homomorphic multiplication depth. Experimental results demonstrate that NetAdapt delivers $6.68\times$ to $12.92\times$ improvements in inference running time compared to state-of-the-art solutions, allowing network-agnostic PPML.