International Association
for Cryptologic Research

Privacy-preserving machine learning (PPML) is critical for protecting sensitive data in domains like healthcare, finance, and recommendation systems. Fully Homomorphic Encryption (FHE) and Secure Multi-Party Computation (MPC) are key enablers of secure computation, yet existing hybrid approaches often suffer from fixed protocol assignments, resulting in inefficiencies across diverse network environments, such as LANs and WANs. To address this, we introduce CostSphere, a cost-model-driven framework that dynamically assigns FHE and MPC protocols to optimize computational efficiency under varying network conditions. Utilizing a predictive cost model based on MLIR’s TOSA-level dialect and an ILP-based solver, CostSphere ensures robust performance for Transformer-based models. Experimental results demonstrate that CostSphere delivers $6.68\times$ to $12.92\times$ improvements in inference runtime compared to state-of-the-art solutions like BumbleBee (NDSS ’25), enabling scalable and network-agnostic PPML across diverse computational scenarios.