Survey of research on efficient secure inference for deep learning

Secure inference in deep learning with homomorphic encryption and secure multi-party computation has brought substantial computational and communication overhead. To address this challenge, a systematic review of existing research on accelerating secure inference was provided. First, a comprehensive comparison was conducted for the technical approaches and optimization methods of two critical components: security protocols and inference models. For secure protocols, the performance and efficiency of different cryptographic primitive solutions were analyzed by distinguishing between linear and nonlinear operations. For inference models, major optimization strategies were evaluated to balance accuracy and computational efficiency. Furthermore, the costs associated with existing schemes were analyzed, and the core framework of secure inference was outlined, expanding the practical considerations for efficient secure inference. Finally, by discussing the challenges faced by secure inference, potential future research directions aligned with practical application needs are proposed.