SPU: Secure Processing Unit

SPU (Secure Processing Unit) aims to be a provable, measurable secure computation device, which provides computation ability while keeping your private data protected.

Project status

Currently, we mainly focus on provable security. It contains a secure runtime that evaluates XLA-like tensor operations, which use MPC as the underline evaluation engine to protect privacy information.

Build

Prerequisite

Docker

## start container
docker run -d -it --name spu-gcc11-anolis-dev-$(whoami) \
         --mount type=bind,source="$(pwd)",target=/home/admin/dev/ \
         -w /home/admin/dev \
         --cap-add=SYS_PTRACE --security-opt seccomp=unconfined \
         --cap-add=NET_ADMIN \
         --privileged=true \
         registry.hub.docker.com/secretflow/spu-gcc11-anolis-dev:latest

# attach to build container
docker exec -it spu-gcc11-anolis-dev-$(whoami) bash

Linux

Install gcc>=11.2, cmake>=3.18, ninja, nasm>=2.15, python==3.8, bazel==5.1.1

python3 -m pip install -r docker/requirements.txt

macOS

# macOS >= 11, Xcode >=13.0

# Install Xcode
https://apps.apple.com/us/app/xcode/id497799835?mt=12

# Select Xcode toolchain version
sudo xcode-select -s /Applications/Xcode.app/Contents/Developer

# Install homebrew
https://brew.sh/

# Install dependencies
brew install bazel cmake ninja nasm

# Install python dependencies
python3 -m pip install -r docker/requirements.txt

Build & UnitTest

# build as debug
bazel build //... -c dbg

# build as release
bazel build //... -c opt

# test
bazel test //...

# [optional] build & test with ASAN or UBSAN, for macOS users please use configs with macOS prefix
bazel test //... --config=[macos-]asan
bazel test //... --config=[macos-]ubsan

Build docs

# prerequisite
pip install -U sphinx

cd docs & make html  # html docs will be in docs/_build/html

FAQ

How can I use SPU?

SPU could be treated as a programmable device, it's not designed to be used directly. Normally we use SecretFlow framework, which use SPU as the underline secure computing device.

SPU python package also contains a simple distributed module to demo SPU usage, but it's NOT designed for production due to system security and performance concerns, please DO NOT use it directly in production.

Acknowledgement

We thank the significant contributions made by Alibaba Gemini Lab.

skyer2/spu