A semi-honest three-party secure and efficient neural network privacy-preserving inference framework

Neural networks have demonstrated remarkable potential in a multitude of domains, such as image recognition, speech recognition, natural language processing, and medical diagnosis. However, with their widespread application, privacy concerns regarding user data have become increasingly prominent. During the process, neural network models may involve vast amounts of sensitive personal information. The critical research question now lies in how to harness the value of data through neural networks while safeguarding users' privacy from being compromised. This paper presents a semi-honest three-party privacy-preserving inference framework for neural networks, designed to conduct neural network inference while ensuring the protection of user data from being disclosed. The framework effectively shifts the computational overhead to the preprocessing stage. Within this framework, our multiplicative protocol requires only 2l bits for online computations, significantly reducing communication complexity and runtime, thus enabling both efficient and secure model inference.