secretflow/spu

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

What it solves

SPU (Secure Processing Unit) provides a way to perform computations on private data while keeping that data protected. It acts as a programmable device for secure computation, ensuring that sensitive information remains private during processing.

How it works

SPU utilizes a secure runtime that evaluates XLA-like tensor operations. It uses Secure Multi-Party Computation (MPC) as its underlying evaluation engine to ensure that privacy is provably protected during the computation process.

Who it’s for

It is primarily designed to be used as an underlying secure computing device within the SecretFlow framework, rather than as a standalone tool for end-users.

Highlights

  • Provable Security: Focuses on providing a measurable and provable level of security for data privacy.
  • XLA-like Operations: Supports tensor operations similar to XLA for efficient computation.
  • MPC-powered: Uses Secure Multi-Party Computation as the core engine for the privacy-preserving mechanism.
  • Hardware Support: Compatible with various platforms including Linux (x86_64 and aarch64), macOS, and experimental support for NVIDIA GPUs via CUDA.

Related

  • Project
    BUAA-BDA/OpenHuFu

    An open-source data federation system for efficient and secure query processing using secure multi-party computation (SMC) to prevent raw data sharing.

  • Project
    tensorflow/tensorflow

    An end-to-end open source platform for machine learning that provides a flexible ecosystem of tools and libraries for building and deploying ML-powered applications.

  • Project
    OpenMined/TenSEAL

    A library for performing homomorphic encryption operations on tensors, enabling secure computation on encrypted data without decryption.

  • Project
    Dstack-TEE/dstack

    An open framework for confidential AI that allows deploying AI applications in hardware-isolated confidential VMs with cryptographic privacy guarantees.

  • Project
    ASCIT31/Dark-Moon

    DarkMoon is an open‑source, self‑hosted platform that uses a fleet of AI agents to run fully automated penetration tests. It orchestrates 50+ real offensive tools via a secure “MCP” gatekeeper, masks all sensitive data before it reaches the language model, and returns reproducible proof for every vulnerability. Supports web, cloud, AD, Kubernetes, databases, IoT and AI/LLM endpoints, runs on local LLMs for privacy, and is GPL‑3.0 licensed.