MPC ModelsPrivate collaborative AI Training
Private collaborative AI Training

MPC Models

Three collaborators train one logistic-regression model over private rows. Arcium runs the MPC computation; Solana devnet records the hashes, accounts, and signatures an auditor needs to inspect the receipt.

3
Private parties
private
1
Rows per party
private
13/28
Features trained/output
public
0
Public source rows
private
Private
Encrypted compute
Public
Aegis IndustriesUS-CANimbus CapitalSGValor HealthCH-ZHMPCencryptedcomputeTrained model28 weights, 1 biasSolana devnetaudit logRAW ROWS NEVER LEAVESHARES ARE COMBINEDANYONE CAN VERIFY

How a training run happens.

The protocol does one thing: combine multiple private inputs into one model, and emit enough on-chain evidence that an outsider can inspect the receipt.

Step 1

Private datasets stay local

Each collaborator keeps source rows in its own environment. The public never sees raw rows, names, labels, or feature values.

Step 2

They publish commitments

A commitment is a signed fingerprint of a dataset. It proves which private input was used — without exposing the input itself.

Step 3

MPC trains over encrypted inputs

The network jointly trains the model across private inputs. The computation uses the combined signal; individual rows stay hidden.

Step 4

The trained model is verifiable

The final receipt stores linked accounts, signatures, and the public model commitment on devnet so anyone can rerun the checks later.

What each role can actually see.

Six artifacts move through the run. Switching role recolors every card by what the selected role can access; the artifact values come from the featured public receipt.

NOT ACCESSIBLE

Raw training rows

Classified as private

Lives only in each owner's environment.

NOT ACCESSIBLE

You never see these rows. They are not on Solana and are not in the model file.

NOT ACCESSIBLE

Sensitive feature values & labels

Classified as private

Stays in the owner's environment, normalized before sharing.

NOT ACCESSIBLE

Hidden behind the privacy boundary. Not present in the public receipt.

VISIBLE TO YOU

Signed dataset commitment

Classified as public

Owner publishes a hash before training.

Aegis Industries Inc. published a signed commitment without exposing source rows.

VISIBLE TO YOU

Agreed data schema

Classified as public

All parties agree on the column shape before training.

The public receipt pins the schema hash for verifier checks.

VISIBLE TO YOU

Training circuit hash

Classified as public

Hash of the exact MPC training program that was run.

Historical receipt from before the mpcmodels:* rename.

VISIBLE TO YOU

The trained model itself

Classified as public

Revealed when the MPC run finalizes.

The verifier recomputes the model hash from the revealed weights and bias.

Visible to this roleEncrypted shares onlyNot accessible
Card color reflects what the selected role can access, not the artifact's classification.

What lives on Solana devnet.

The chain holds the structure needed to re-run the proof: who joined, what schema they agreed to, which MPC circuit was used, and which public model came back. Hover or tap any field to read what it commits to.

solana devnet - TrainingJobAccount
Open account on Orb
3T43E8zJbevm1Nj7mLvwuZPErdyMEvm8bVax5Lv9NFme
schemadata/schemas/threats-v1.json
training_circuittrain_logreg_leaf13
program_accountswrapper, training, registries, Arcium comp-def
participants3 private parties
dataset_commitments3 signed commitments recorded
included_bits0b111 (7)
model_output13 trained features + 1 bias
finalized_atMay 27, 2026, 11:41 PM UTC
Receipt-backed field reader
job_account

What this field tells you.

The Solana devnet account that stores this training receipt. The gallery links to this account directly so you can read the raw state.

You can do this yourself

Read the devnet account with solana account <training-job-account> --url devnet, then compare it with the public receipt data shown here.