dropPINdropPIN
Launch App
DePIN · Decentralized AI · Federated Learning

Train together.
Share nothing.

A federated learning protocol on Solana. Nodes train locally and drop their model update on-chain. Fund a model in SOL or $dPIN — rewards flow to honest contributors, junk gets slashed.

Order a modelHow it works
compute
off-chain
trust & rewards
on-chain
funded in
SOL / $dPIN
train locallyshare nothingdrop the PINrobust aggregationstake · slashSOL → $dPINon-chain rewardssoulbound reputationdifferential privacyFedAvgbuy & burnno central servertrain locallyshare nothingdrop the PINrobust aggregationstake · slashSOL → $dPINon-chain rewardssoulbound reputationdifferential privacyFedAvgbuy & burnno central server
The problem

Federated learning works in theory. It stalls in practice.

Two failures kill it every time — and a third makes it unsafe. dropPIN removes all three by moving coordination, incentives and accountability on-chain.

01

No trusted coordinator

Classic FL needs a central server to run rounds and aggregate — a single point of trust and failure.

02

No incentive to contribute

Data owners carry the compute cost and privacy risk but get nothing back. Supply never shows up.

03

No defense against bad actors

Anyone can submit junk or a poisoning update. Without stake and verification, the model quietly rots.

How it works

The lifecycle of a round

Heavy artifacts (weights, deltas) live in decentralized storage. Only hashes, commits and metadata touch the chain.

  1. 01

    Open round

    An Anchor program opens a round and pins the global model hash on-chain.

  2. 02

    Stake & register

    Participants stake $dPIN as skin in the game and join the round.

  3. 03

    Train · drop the PIN

    Nodes train on private data, publish the delta to storage, commit the hash.

  4. 04

    Validate

    Validation set, robust aggregation or committee vote catch junk and poisoning.

  5. 05

    Reward · slash

    A new model ships. Honest work earns $dPIN; malicious stake is slashed.

on-chain: round registry · stake escrow · update hashes · validation · reward distribution · reputation
Order · Swap

Pay in SOL.
Buy pressure lands on $dPIN.

Fund a model in SOL today — the order is recorded on-chain and its reward pool is denominated in $dPIN at launch. When the token goes live, funding routes SOL → $dPIN on a DEX, creating real, usage-driven buy pressure.

  • Buy pressure

    Each round swaps SOL → $dPIN on a DEX — recurring, usage-driven demand.

  • Locked supply

    Participants must stake $dPIN to take part, removing supply from circulation.

  • Deflation

    A protocol fee on every round buys back and burns $dPIN over time.

note: on a thin pool, in-tx swaps risk slippage/MEV — early rounds batch swaps through the treasury (TWAP) instead of swapping at settle.

Fund a training round

preview
Reward pool (escrowed)balance
SOL
≈ $15.24
≈ 386 $dPIN reward pool at TGE · 34 burned (8%)

Model / task

Max nodes

node stake 0.02 SOL each

8
Launch app to order

real on-chain round · SOL escrowed in the coordinator program

Skin in the game

Stake to participate. Get slashed for harm.

A refundable $dPIN stake makes spam and poisoning more expensive than any reward they could earn. Reputation compounds the effect.

Honest update

reward
  • Improves the model on the validation set → passes.
  • Earns $dPIN weighted by measured contribution, not by showing up.
  • Builds soulbound reputation → higher vote weight and reward share.

Junk / poisoning

slash
  • ✗Robust aggregation (Krum / trimmed mean) filters outliers first.
  • ✗Updates that degrade the model get their stake partially slashed.
  • ✗Forging usefulness costs more than the reward — spam stops paying.
$dPIN utility

Three sources of demand — reinforcing each other

The token is required to use the network, not bolted on. Demand, locked supply and deflation compound.

Usage-driven demand

Customers fund rounds in SOL, which the protocol converts into $dPIN — recurring buy pressure tied to real model training, not emission alone.

SOL$dPIN$dPIN/ round

Stake & slash

Participation requires a $dPIN stake. Honest work is rewarded; harmful updates are slashed — locking supply and aligning behaviour.

Reputation & governance

Soulbound reputation (cNFT) plus staked $dPIN weight committee votes and reward shares for proven contributors.

Roadmap

Honest about what's built and what's research

The MVP proves a narrow loop: honest update → validation → reward, junk → slash. Privacy and trust-minimization layer on after — labeled, not overpromised.

now

MVP — prove the loop

  • One simple model, 5–10 nodes, synchronous FedAvg
  • Validation set + slashing (no ZK)
  • Anchor programs · web3.js · Flower for training
  • Storage on Arweave / IPFS
next

Harden incentives & privacy

  • SOL → $dPIN swap routing + buy-and-burn treasury
  • Differential privacy in the training client
  • Committee validation with on-chain voting
  • Soulbound reputation (cNFT)
research

Trust-minimized frontier

  • Secure aggregation (MPC) — remove aggregator centralization
  • ZK proof-of-training — experimental, not production-ready
  • Decentralized / rotating aggregators
Get involved

Bring your data. Keep it private.
Earn for the model you help build.

Run a node, fund a model in SOL, or read the protocol design.

Order a modelRead the docs