FPGA-accelerated secure machine learning

Multi-party computation on FPGA SmartNICs with RDMA

Multi-party computation (MPC) lets several parties train or evaluate a model jointly without revealing their inputs, but every multiplication needs a round of communication between parties. This project runs the MPC engine on FPGA SmartNICs and connects them with the Coyote RDMA stack, so secret-shared data moves between parties without touching the host CPU and communication overlaps with computation. Paper (IEEE Xplore)

FPGA-accelerated multi-party computation Three parties each pair a host with an FPGA SmartNIC that runs the MPC engine. The SmartNICs exchange secret-shared data over RDMA through a 100 Gb/s switch, bypassing the CPU. Party 1 Host CPU FPGA SmartNIC Party 2 Host CPU FPGA SmartNIC Party 3 Host CPU FPGA SmartNIC 100 Gb/s switch · RDMA, CPU bypass Secret-shared data only: no party sees another's inputs
2.1×
faster convolution layers than existing MPC frameworks

Current work extends this to GPU–FPGA acceleration with GPUDirect RDMA, aiming to cut host-memory copies from the secure-training data path.

References

2025

  1. Accelerating Multi-Party Computation Using Heterogeneous Systems
    Outstanding Student Paper Award
    Xiteng Yao, Shining Yang, Mayank Varia, and Martin Herbordt
    In 2025 IEEE High Performance Extreme Computing Conference (HPEC), 2025