Maliciously Secure Privacy-Preserving Neural Network Inference Framework Based on Function Secret Sharing
Mengbo Li, Rui Jiang · 2025
With the widespread adoption of machine learning as a service (MLaaS), ensuring the privacy of user data and neural network models has become a critical challenge. Current privacy-preserving neural network inference protocols predominantly follow semi-honest security assumptions, leaving practical and efficient solutions secure against malicious adversaries an open and pressing issue. In this paper, we propose MSPI-FSS, a maliciously secure privacy-preserving neural network inference framework based on function secret sharing (FSS), designed explicitly for secure client-server inference scenarios. To robustly counteract malicious adversaries, we integrate additive secret sharing with SPDZ message authentication code (MAC) to ensure both input privacy and inference result integrity. Furthermore, we design a secure two-party comparison protocol for non-linear layers based on FSS, which significantly reduces communication rounds. Experimental evaluations demonstrate that our proposed MSPI-FSS framework achieves significant performance improvements compared to existing maliciously secure protocols, such as Muse.