Secretflow

Added

- Component: Suppport Early Stop in SGB via XGBoost Callbacks.
- Component: Add storage interface and oss/s3 storage impl.
- FLModel: Add new sample methods for datasets.
- Device Support cheetah h2a.
- Tuner optimizes resource allocation logic.
- Docs: Add SecretFlow Benchmark Results.
- Docs: Add more docs for SS-GLM.

Changed

- Use ray instead of secretflow-ray and bump to ray-2.9.

Added

- Add Criteo(100M) Dataset.
- Add split learning nn component.
- Add sgb & ss_xgb model export.
- Split learning support prediction for file input.
- GPU support for split learning attack algorithm.

Changed

- Unify nn Callback naming.

Added

- Split learning in PyTorch supports GPU in debug mode.
- Split learning in PyTorch supports pipeline strategy.
- Sparse compression supports the compression of multi-dimensional data, such as images.
- SLModel supports input in the form of file names and allows for multiple label inputs.

- Component: prediction operator results can now save feature columns.
- Component: binning operators can now view reports directly.
- Component: binning rule modification is supported.
- Component: support model exports, support both online and offline prediction.

Fixed

- fix: bump rayfed to 0.2.0a18 to fix shutdown lock.
- Component: improve GLM training accuracy.
- Component: reduce GLM training memory and computation cost.

Added

- Add a new feature of chunked computation to control the memory peak.
- The binning component supports displaying binning rules.

Fixed

- Fix a bug that occurs when saving the SGB model fails when the number of trees in the model is 1.
- Fix bin rule report.

Added

- sgb/ss_xgb/ss_glm/ss_sgd can save specified columns with result now.
- ss_glm: expose two params infeed size and newton.
- SplitLearning
- Add a defense method against data reconstruction attacks.
- Adapt replay attack and gradreplace attack for deepfm.
- Add expoit attack.

Changed

- Component: update fillna descriptions.

Fixed

- Fix SS-GLM toturial.

Added

- Add grad replace attack and replay attack.
- Add autoattack benchmark examples.
- Expose job_name param in sf.init.

Changed

- Bump rayfed version: optimizing Error Propagation and Capture.
- Component: woe_bins requires at least 5 bins to read.
- Component: add barrier_on_shutdown as sf cluster config.