Reproduce “LRD-MPC: Efficient MPC Inference through Low-rank Decomposition” on OpenResearch

About this paper

Tingting Tang, Yongqin Wang, Murali Annavaram

The LRD-MPC framework improves the efficiency of secure multi-party computation (MPC) for machine learning inference by combining low-rank decomposition with truncation skipping and efficient linear layer concatenation. This approach yields up to 33% faster online inference, up to 52% lower GPU energy consumption, and substantial reductions in offline phase costs across various deep learning models and MPC protocols, while maintaining model accuracy.

Key insights

• MPC-specific optimizations are crucial when applying low-rank decomposition to linear layers, as naive application introduces new overheads that undermine efficiency gains.
• Strategic skipping of intermediate truncation steps in concatenated linear operations within fixed-point MPC can save significant communication and computation without compromising overall accuracy.
• Pipelining and overlapping communication with computation during the execution of sequentially decomposed linear layers effectively masks network latency, improving online phase performance.

Summary by alphaXiv.