Privacy-Preserving Floating-Point Operations: Optimizing Division in 2PC Frameworks

This paper proposes an optimized floating-point division protocol to address the issues of high communication overhead and computational complexity in existing protocols, while ensuring accuracy and security. The proposed method improves the mantissa computation process, significantly reducing data transmission volume and interaction rounds. Experimental results demonstrate that the proposed protocol outperforms existing mainstream methods in terms of runtime, communication overhead, and efficiency, thus optimizes complex nonlinear functions such as Sigmoid, Tanh, and Softmax. This research provides new insights into optimizing floating-point operations and establishes a solid foundation for the application of privacy-preserving computation in high-precision scenarios.