SecureRouter Framework Accelerates Encrypted Neural Network Inference with Adaptive Model Selection

SecureRouter has unveiled a novel encrypted routing and inference framework aimed at enhancing the speed of secure transformer inference by adaptively selecting models based on encrypted data. This innovation tackles the shortcomings of current privacy-preserving inference techniques that depend on Secure Multi-Party Computation (MPC). Traditionally, these methods utilize a single, static transformer model for all encrypted data, leading to inefficiencies as different inputs necessitate various model sizes to optimize accuracy and computational expense. By creating a cohesive encrypted pipeline that combines a secure router with a model pool optimized for MPC, SecureRouter facilitates coordinated routing, inference, and protocol execution while ensuring cryptographic security. This strategy seeks to address the slow and expensive nature of previous systems that have impeded the practical use of secure neural network inference. The methodology allows cloud servers to handle client inputs without decryption, thus safeguarding privacy during computation. This research was published on arXiv under the identifier 2604.15499v1, classified as a cross announcement type.