MPCMamba: Privacy-preserving inference for Mamba models via secure multi-party computation
Yongqiang Yu, Yuliang Lu, Xuehu Yan, Wei Yan, Shengyang Luo · Applied Soft Computing · 2025
Deep learning is increasingly critical across high-stakes domains, but privacy risks during model inference remain a major concern. Secure multi-party computation (SMPC) offers a promising solution by enabling privacy-preserving collaborative inference without exposing sensitive data. The recently proposed Mamba model, which outperforms Transformer in certain tasks, presents unique challenges for SMPC due to its state-space architecture and nonlinear operations. This paper introduces a framework for executing Mamba model inference under SMPC while preserving privacy. The framework natively supports linear operations and securely computes nonlinear functions – including square roots, exponentials, logarithms, and SiLU activations – without altering the original model architecture. Experimental results demonstrate that the proposed method achieves accuracy improvements of 2.4%, 2.84%, and 8.23% on CIFAR-10, CIFAR-100, and Tiny-ImageNet datasets, respectively, compared to existing SMPC-based vision Transformer approaches, MPCViT. Additionally, inference latency is reduced by factors of 2.14 × , 2.14 × , and 26.13 × on these benchmarks, significantly advancing efficient and secure deployment of state-space models in privacy-sensitive scenarios.