Skip to content

SecretFlow-SPU: A Performant and User-Friendly Framework for Privacy-Preserving Machine Learning

Authors: Ma, Zheng, Fang et al.
Year: 2023
Venue: USENIX ATC 2023

type: system collection: catalogue status: auto

Links: link

Summary

Auto-extracted summary (unreviewed)

SecretFlow-SPU is an MPC-based PPML framework consisting of a frontend compiler that lowers ML programs (via HLO) into an MPC-specific MLIR dialect (PPHLO) and a C++ backend runtime that executes the optimized IR with semi-honest MPC protocols (ABY3, SPDZ2k, Cheetah). This lets existing JAX/TensorFlow/PyTorch programs run privately with minor modifications. In private training of four MNIST neural networks under 3PC, SPU is fastest in 23 of 24 settings, up to 4.1x faster than MP-SPDZ and 2.3x faster than TF Encrypted under WAN.

Abstract

With the increasing public attention to data security and privacy protection, privacy-preserving machine learning (PPML) has become a research hotspot in recent years. Secure multiparty computation (MPC) that allows multiple parties to jointly compute a function without leaking sensitive data provides a feasible solution to PPML. However, developing efficient PPML programs with MPC techniques is a great challenge for users without cryptography backgrounds. Existing solutions require users to make efforts to port machine learning (ML) programs by mechanically replacing APIs with PPML versions or rewriting the entire program. Different from the existing works, we propose SecretFlow-SPU, a performant and user-friendly PPML framework compatible with existing ML programs. SecretFlow-SPU consists of a frontend compiler and a backend runtime. The frontend compiler accepts an ML program as input and converts it into an MPC-specific intermediate representation. After a series of delicate code optimizations, programs will be executed by a performant backend runtime as MPC protocols. Based on SecretFlow-SPU, we can run ML programs of different frameworks with minor modifications in a privacy-preserving manner. We evaluate SecretFlow-SPU with state-of-the-art MPCenabled PPML frameworks on a series of ML training tasks. SecretFlow-SPU outperforms these works for almost all experimental settings (23 out of 24). Especially under the wide area network, SecretFlow-SPU is up to 4.1× faster than MP- SPDZ and up to 2.3× faster than TF Encrypted.
Source: pdf

Extracted details

Field Value Source Confidence
contribution Frontend compiler converting ML programs into the PPHLO MPC-specific intermediate representation with compiler optimizations llm 0.90
contribution Performant C++ backend runtime executing PPHLO as MPC protocols (ABY3, SPDZ2k, Cheetah) llm 0.90
contribution Compatibility with mainstream ML frameworks via Python APIs so programs need only minor modifications llm 0.90
contribution Open-sourced framework and evaluation against MP-SPDZ, TF Encrypted and CrypTen on private training tasks llm 0.90
schemes secret sharing llm 0.90
schemes replicated secret sharing (ABY3) llm 0.90
schemes SPDZ2k llm 0.90
schemes Cheetah llm 0.90
libraries SecretFlow-SPU llm 0.90
libraries MLIR llm 0.90
libraries JAX llm 0.90
libraries bRPC llm 0.90
libraries MP-SPDZ llm 0.90
libraries TF Encrypted llm 0.90
libraries CrypTen llm 0.90
models four neural networks A-D (Wagh et al. MNIST models) llm 0.90
datasets MNIST llm 0.90
tasks privacy-preserving neural network training (image classification) llm 0.90
hardware 3x Alibaba Cloud ecs.g7.xlarge instances (4 vCPU, 16 GB RAM); LAN 10.1 Gbps/0.1 ms, WAN 300 Mbps/40 ms llm 0.90
threat model semi-honest (3PC ABY3, N-party SPDZ2k, 2PC Cheetah) llm 0.90
interaction pure MPC llm 0.90
code url https://github.com/secretflow/spu.The (URL found in the full text) regex 0.50
code url https://github.com/secretflow/spu/blob/main/docs/reference/pphlo_op_doc.md (URL found in the full text) regex 0.50
code url https://github.com/apache/incubator-brpc (URL found in the full text) regex 0.50
code url https://github.com/secretflow/spu llm 0.90
llm relevance medium llm 0.90
key result settings where SPU is fastest: 23 of 24 (MNIST training vs MP-SPDZ, TF Encrypted, CrypTen) llm 0.90
key result training speedup over MP-SPDZ: up to 4.1x (WAN, semi-honest 3PC) llm 0.90
key result training speedup over TF Encrypted: up to 2.3x (WAN, semi-honest 3PC) llm 0.90
key result training speedup over TF Encrypted: 1.4-4.6x (LAN) llm 0.90
key result code base size: >50k LOC C++ and 3k LOC Python (SPU implementation) llm 0.90

System card: SecretFlow-SPU

Group 5. Interactive MPC and HE+MPC hybrids
Family MPC compiler (ABY3, Cheetah, SEMI2K)
Arithmetic Fixed point
Interaction Interactive 2PC/3PC
Model scale LLaMA-7B (PUMA)
Status Widely used
Relation to the SoK JAX-to-MPC compiler used by PUMA, BumbleBee, Ditto, Nimbus.
Note Ant Group open-source stack. Comparable role to OpenFHE for the MPC world.
Code https://github.com/secretflow/spu
Systems page SecretFlow-SPU

Keyword mentions

Counts of vocabulary terms in the full text (first page of occurrence in brackets).

  • scheme: MPC 119 [p.2], Secret sharing 16 [p.4], TEE / confidential computing 3 [p.13]
  • library: SecretFlow-SPU 199 [p.1], MP-SPDZ 18 [p.2], CrypTen 17 [p.2], PyTorch 16 [p.2], TensorFlow 12 [p.2], ABY / ABY3 4 [p.2]
  • model: RNN / LSTM 6 [p.11], Autoencoder 4 [p.3], CNN 2 [p.8], LeNet 1 [p.17], Logistic regression 1 [p.12], Decision tree 1 [p.11]
  • dataset: MNIST 6 [p.3], Breast cancer (Wisconsin) 1 [p.12]
  • hardware: GPU 3 [p.4], TPU 2 [p.4], Intel Xeon 1 [p.18]
  • technique: ReLU 20 [p.17], Compiler 18 [p.2], Matrix multiplication 11 [p.4], Convolution 11 [p.4], Three-party computation 7 [p.10], Semi-honest model 7 [p.5], Two-party computation 5 [p.10], Encrypted training 4 [p.10], Federated learning 3 [p.15]
  • topic: Privacy-preserving machine learning 30 [p.2], Open source 9 [p.3]

Files

Path (under papers/) Collection Pages Size SHA-256 Note
references/catalogue/5-079_secretflow-spu-a-performant-and-user-friendly-framework-for.pdf catalogue 18 2.8 MB 33a32d663023

Database key: secretflow-spu-a-performant-and-user-friendly-framework-for-privacy · created 2026-11-05 · updated 2026-11-05 · edit with python3 tools/paperdb.py set secretflow-spu-a-performant-and-user-friendly-framework-for-privacy COLUMN VALUE