Privacy-Preserving Multi-Party Machine Learning Based on Trusted Execution Environment and GPU Accelerator
Haibo Tian, Shuangyin Ren, Chao Chen · 2024
In multi-party machine learning scenarios involving sensitive data, safeguarding user and model privacy holds utmost importance. Current approaches often disclose model parameters to expedite training, but this poses significant privacy concerns. This paper introduces MOML (Matrix Outsourcing for multi-party Machine Learning), a framework for multi-party learning that relies on outsourced matrix multiplication. MOML harnesses the power of both TEE (Trusted execution environment) and GPU (Graphics Processing Unit), effectively offloading inefficient matrix operations from TEE to GPU, thereby enhancing training efficiency without compromising privacy. At the heart of MOML lies an enhanced matrix multiplication outsourcing algorithm, tailored for GPU computing and integrity checks. Rigorous tests conducted on the MNIST and CIFAR-10 datasets reveal remarkable performance improvements with MOML.