Interactive Tutorial
Try SecureRoute
Run SecureRoute directly in your browser via Google Colab. The Tutorial walks you through the encrypted routing and inference pipeline step-by-step with runnable examples.
A single sentiment classification query flows through SecureRoute: the sentence is split into secret shares on the client, scored by the secure router, dispatched to the cheapest sufficient model, and reconstructed back on the client. Toggle Easy and Hard to see cost-aware routing in action.
Existing privacy-preserving inference frameworks such as MPCFormer and SecFormer execute a single, fixed transformer model under Secure Multi-Party Computation (MPC) for all inputs, making them inherently slow and computationally expensive, which restricts their real-world adoption.
We present SecureRoute, an end-to-end encrypted routing and inference framework that accelerates secure transformer inference through input-adaptive model selection under encryption. At the framework level, SecureRoute establishes a unified encrypted execution pipeline that tightly integrates a secure router with a model pool, enabling seamless coordination between routing, inference, and MPC protocol execution while preserving complete data and model confidentiality. At the algorithmic level, SecureRoute introduces an MPC-cost-aware secure router that predicts per-model inference cost and utility from encrypted input statistics, and an MPC-optimized model pool whose architectures and quantization schemes are co-trained to minimize MPC communication and computation overhead for router decisions.
Built atop state-of-the-art SMPC protocols, SecureRoute achieves a latency reduction by 1.95Γ with negligible accuracy loss, offering a practical path toward scalable and efficient secure AI inference.
SecureRoute selects the optimal transformer model per-input under encryption, breaking free from the one-size-fits-all paradigm of prior MPC inference systems.
A unified encrypted pipeline integrates the secure router with an MPC-optimized model pool, preserving full data and model confidentiality throughout routing and inference.
The router and model pool are co-trained with MPC cost awareness, jointly optimizing for communication overhead, computation cost, and inference accuracy.
Achieves a 1.95Γ latency reduction over prior work with negligible accuracy loss, offering a practical path toward scalable and efficient secure AI inference.
Model architectures and quantization schemes are co-designed to minimize MPC communication and computation overhead, enabling efficient encrypted inference at scale.
Deep integration of routing decisions with MPC protocol execution ensures that efficiency gains are realized end-to-end in the cryptographic setting.
Run SecureRoute directly in your browser via Google Colab. The Tutorial walks you through the encrypted routing and inference pipeline step-by-step with runnable examples.
If you find SecureRoute useful in your research, please cite our paper:
@inproceedings{zhang2026secureroute,
title = {SecureRoute: Input-Adaptive Encrypted Routing for Efficient
and Accurate MPC-Based Transformer Inference},
author = {Zhang, Yukuan and Zheng, Mengxin and Lou, Qian},
booktitle = {Proceedings of the 63rd ACM/IEEE Design Automation Conference (DAC)},
year = {2026}
}