Machine learning for data you're not allowed to see

What it does

SecretFlow is a Python framework for training models and analyzing data that is split across organizations. It abstracts secure multiparty computation, homomorphic encryption, federated learning, and trusted execution environments into plain and “secret” device layers. A device-flow layer sits above that, modeling algorithms as DAGs of device objects, while the top algorithm and workflow layers handle partitioned data, model training, and hyperparameter tuning.

The interesting bit

Rather than forcing you to commit to one cryptographic technique, SecretFlow lets you compose workflows from interchangeable devices—MPC, HE, TEE, or plain compute—and executes them through backends like SPU and HEU. The framework essentially turns gnarly cross-party cryptography into a dataflow problem.

Key highlights

• Unifies multiple privacy-enhancing technologies behind one device abstraction
• Supports both horizontal and vertical data partitioning
• Ships with companion projects: SPU for secure computation, HEU for homomorphic encryption, SCQL for private SQL analytics, and Kuscia for K3s-based task orchestration
• Explicitly prohibits production use of non-release versions

Caveats

• The README is architecture-heavy and light on concrete API examples or benchmarks
• Non-release builds are explicitly disclaimed for production use

Verdict

Useful if you are building cross-organizational ML pipelines and want one Python stack instead of a pile of incompatible cryptographic libraries. If your data stays inside one trust boundary, it is overkill.