dropPIN docs

Classic federated learning stalls for two reasons: it needs a trusted central coordinator, and data owners have no incentive to contribute. dropPIN removes both by moving coordination, rewards and accountability on-chain.

Heavy compute (training, aggregation) stays off-chain. The chain only stores hashes, commits and metadata, and holds the reward pool and stakes in escrow.

• Cheap, frequent transactions — FL has many rounds, commits and votes.
• Anchor programs map cleanly to “round → collect → aggregate → pay”.
• Native token for incentives; cNFTs for soulbound reputation.
• Fast finality for synchronous rounds.

1. A customer opens a round and escrows the reward pool.
2. Nodes stake $dPIN (skin in the game) and register.
3. Local training off-chain → publish the update → commit its hash on-chain (“drop the PIN”).
4. Validation: a holdout set, robust aggregation, or a committee vote.
5. A new global model ships; honest work earns rewards, malicious stake is slashed.

On-chain (Anchor): round registry, staking escrow, update-hash commits, validation verdicts, reward distribution and reputation.

Off-chain: each node trains on its private data and publishes the weight delta to decentralized storage (Walrus / Arweave / IPFS). Only the hash touches the chain.

Aggregator: collects updates and runs robust aggregation (FedAvg + outlier trimming); rotatable or run via secure aggregation (MPC) on the roadmap.

The hard problem of any “FL on-chain” design — graded from simple to ambitious:

1. Validation set — accept an update only if it improves a holdout metric.
2. Robust aggregation (Krum, trimmed mean) — statistically reject poisoning.
3. Committee validation — randomly chosen nodes vote on-chain.
4. ZK proof-of-training — the crypto-native ideal; experimental and costly today.

Gradients can leak data. Production privacy needs differential privacy or secure aggregation (MPC). These are on the research roadmap and not enabled in the MVP — the current build proves the incentive loop, not privacy guarantees.

Three reinforcing sources of demand:

• Usage demand — customers fund rounds; funding routes SOL → $dPIN, creating buy pressure tied to real training.
• Locked supply — participation requires staking $dPIN.
• Deflation — a protocol fee on every round buys back & burns $dPIN.

The token is not live yet. Until TGE, rounds are funded and escrowed directly in SOL on mainnet.

A refundable stake makes spam and poisoning more expensive than any reward they could earn. Honest updates earn $dPIN weighted by measured contribution and build soulbound reputation; updates that degrade the model are partially slashed.

The coordinator is a live Anchor program on mainnet.

Accounts: Protocol (PDA), Round (PDA, holds escrow), Participant (PDA). Instructions:

view program on Solana Explorer

• Now — coordinator live on mainnet, SOL escrow, real create / join / commit / claim.
• Next — $dPIN TGE, SOL→$dPIN swaps (Jupiter), staking & rewards in $dPIN, committee validation, soulbound reputation.
• Research — differential privacy, secure aggregation (MPC), ZK proof-of-training, audit + mainnet.

$dPIN is a utility token for staking, rewards and governance within the protocol. This documentation describes a concept (v0.1) on mainnet; it is not an offer to sell or solicitation to buy any asset, and not financial advice. The program is unaudited — do not send mainnet funds.