Cybersecurity threat detection: private training
1 row per party, 1 epoch, 13 trained features over a 28-feature threat schema. The participants do not get each other's source data; they get this trained scoring rule. The on-chain accounts below pin it to the recorded job — same schema, same training code, same revealed weights.
Source rows and labels remain inside each participant's environment.
Party signatures, job state, schema hash, and final model hash.
Solana devnet stores the audit trail for this job.
Logistic-regression rule · 28 weights + 1 bias.
What pushes the score up or down.
Each feature has one learned weight. Positive bars push the predicted score up; negative bars push it down. Inputs are normalized into [−1, 1] before MPC, so the magnitudes are directly comparable across columns.
Positive and negative mean direction in the learned score — not a security verdict. Bias term: -1.667e-3.
Every public account and transaction below opens in Orb on devnet. The Arcium rows identify the computation definition and live computation account used by this MPC run.
The Solana devnet account that stores this receipt. Everything else hangs off of it.
The Anchor program that ties the MPC run to its commitments and finalizes the receipt on-chain.
First-party Arcium Explorer page for the wrapper program that queued this computation.
The MPC training circuit that was executed for this job.
The on-chain list of parties allowed to join training runs.
Where each party's signed dataset commitment is stored.
The key used to verify the finalize signature for this job.
Arcium computation definition the MPC nodes executed.
The transaction that queued the MPC computation onto the cluster.
Arcium account that holds the in-flight MPC state for this job.
First-party Arcium Explorer page for this exact MPC computation run.
The transaction the cluster sent back with the finalized result. After this, the receipt is on-chain.