SecretFlow-SPU: A Performant and User-Friendly Framework for Privacy-Preserving...
summary
SecretFlow‑SPU compiles XLA frontend graphs into an MPC‑specific IR and runs them on a high‑performance runtime that supports pluggable MPC protocols; compiler passes reorder multiplications to cut secret truncations and rewrite max‑pool to emit a reusable one‑hot index, while the runtime uses vectorization and streaming tiles to hide latency. Minimal code changes let a JAX GPT‑2 run by secret‑sharing inputs and parameters across parties. Evaluation shows 23 of 24 benchmarks beat prior work, reaching 4.1× faster than MP‑SPDZ and 2.3× faster than TF Encrypted.
AI-written summary · about
chapters (10)
- 0:00SecretFlow-SPU intro
- 2:34Secure Multi-Party Computation basics
- 5:44MPC-based PPML challenges
- 8:55PPML frameworks overview
- 11:25SecretFlow-SPU solution
- 12:50SPU architecture goals
- 15:35GPT-2 inference with SPU
- 18:48SPU compiler optimizations
- 21:33SPU runtime performance
- 22:17Performance evaluation
related talks
20mDeep Dive into Privacy Enhancing Technologies
38mDemystifying Privacy Preserving Computing
18mFLASH: Towards a High-performance Hardware Acceleration Architecture for Cross-silo...
16mPrivacy Preserving AI
16mBalancing Privacy: Leveraging Privacy Budgets for Privacy-Enhancing Technologies
45mMaking Privacy Technology Accessible: Benchmarks and Platforms
25mOpen Source Model Performance Optimization With SGLang
21m"SECRETS ARE LIES, SHARING IS CARING, PRIVACY IS THEFT." - A Dive into Privacy Preserving