MPC ModelsPrivate collaborative AI Training
Protocol overview

Trained models

Each record corresponds to one finalized training job. Open one to inspect the schema, the revealed weights, and the on-chain anchors that bind them together.

WHAT EACH MODEL IS

A logistic-regression rule. 28 weights, 1 bias.

The training run produces one artifact participants actually use: a small set of learned coefficients that can score future records prepared against the same schema. The accounts and signatures on Solana exist to prove this exact rule came from the agreed inputs and the expected training code.

READING A FEATURE BAR

Each feature has one learned weight. Inputs are normalized into [−1, 1] before training, so bar magnitudes are directly comparable.

Positive weight+1.67e−3Zero0.0000Negative weight−1.68e−3
More signal than one dataset has alone

A shared model learns across separate private cohorts while every source dataset stays under local control.

A simple risk-scoring artifact

Each run outputs a logistic-regression rule: feature weights plus one bias that produce a score for the agreed schema.

A thing auditors can inspect

Because the model is revealed and hashed, anyone can check that the published coefficients match the on-chain record.

May 28, 2026, 12:52 AM

AML / suspicious activity detection

1 row per party, 1 epoch, 13 trained features over a 28-feature AML schema.

3 private parties joined without publishing source rows.
13 trained features, 28-weight output, model hash on devnet.
AML / suspicious activity detectionPrivate multi-party trainingtrain_logreg_leaf13
May 28, 2026, 12:35 AM

ICU mortality

Row2 devnet proof: 2 encrypted rows per party, 1 epoch, 4 trained features, 28-weight output shape.

3 private parties joined without publishing source rows.
4 trained features, 28-weight output, model hash on devnet.
ICU mortalityPrivate multi-party trainingtrain_logreg_leaf13
May 27, 2026, 11:41 PM

Cybersecurity threat detection

1 row per party, 1 epoch, 13 trained features over a 28-feature threat schema.

3 private parties joined without publishing source rows.
13 trained features, 28-weight output, model hash on devnet.
Cybersecurity threat detectionPrivate multi-party trainingtrain_logreg_leaf13