VertiMosaic¶
CPU-first vertical federated learning for heterogeneous tabular data.
VertiMosaic is an open-source research framework for vertical federated learning (VFL) on aligned tabular entities. It is designed around protocol correctness, reproducibility, provenance, auditable communication, explicit privacy boundaries, and CPU-friendly reference implementations.
Get started Architecture GitHub
Privacy posture
Raw-feature locality is a design property of VertiMosaic. It is not presented as equivalent to end-to-end cryptographic privacy. Privacy guarantees are scoped to the exact mechanism, protected release, observer, and threat model documented for each experiment.
Why VertiMosaic¶
-
:material-vector-arrange-below:{ .lg .middle } Strict entity alignment
Order-sensitive entity identifiers and digests can reject equal-length but misaligned party partitions before training, validation, or inference.
-
:material-cpu-64-bit:{ .lg .middle } CPU-first research
Reference logistic-regression and histogram-GBDT VFL implementations are designed to reproduce on developer-class CPU environments without requiring a GPU.
-
:material-shield-lock-outline:{ .lg .middle } Explicit privacy boundaries
Optional clipped-Gaussian releases, secure aggregation, additive sharing, PSI, and Paillier primitives are documented without silently upgrading the default protocol's privacy claims.
-
:material-test-tube:{ .lg .middle } Release-grade evidence
Reproducibility bundles, hashes, environment metadata, privacy audits, communication measurements, release assets, and clean-room reproduction tooling make results inspectable.
Current research surface¶
| Area | Current scope |
|---|---|
| Federation | Vertical federated learning |
| Reference models | VFLLogisticRegression, VFLHistGBDT |
| Compute | CPU supported; GPU not required |
| Exact multi-source public linkage | NPI-linked Open Payments + NPPES + CMS Care Compare/provider data |
| Additional linked benchmarks | MovieLens 1M, UCI Credit, authorized local IEEE-CIS |
| Cross-industry benchmark | Public-source, explicitly semi-synthetic linkage |
| Protected research path | PSI alignment + clipped-Gaussian residual releases with scoped accounting |
| External comparison | Deterministic exchange/normalization contract for third-party VFL frameworks |
| Reproduction | PyPI install path, evidence bundles, attestation schema, validation tooling |
Documentation map¶
- Quick Start — install the package and run the first deterministic CPU experiment.
- Architecture — understand parties, models, transports, evaluation, and reporting boundaries.
- Protocol — entity alignment, message flow, missing-party behavior, and protocol invariants.
- Privacy — what the framework protects, what it does not, and how optional backends change the boundary.
- Benchmarks — distinguish exact linkage, vertical partitioning, and semi-synthetic cross-domain experiments.
- Reproduction — produce inspectable evidence bundles and independently validate a release.
- API Reference — primary public classes and research entry points.
Version¶
The documentation targets the current v0.3.x research surface. For immutable artifacts and measured release evidence, use the corresponding GitHub release rather than assuming the main branch and a published tag are identical.