Secure Decentralized Learning with Incentives on UTXO-based Blockchains

Increasing computational and economic costs in training artificial intelligence motivate the development of alternative decentralized training paradigms. Federated learning achieves decentralization by creating the final model combining local models trained on edge devices. This process does not require sharing local private data with the central service. The blockchain can be used to distribute and collect edge models. The benefits of relying on the blockchain are a native way to provide economic incentives to the local devices and the ability to log events and track training steps. The former incentivizes honest behavior, while the latter enables auditing and testing of the model. We present a peer-to-peer cloud protocol to outsource computations using an unspent transaction output blockchain relying on proof of work. In designing this protocol, we face three main challenges: how to communicate the computation to the computing party, how to make its execution safe for them, and how to ensure the computing party executed it faithfully. Our solution introduces a verification script in the outsourced computation. The computing party receives the computation as an obfuscated executable and provides the verification output to show they faithfully executed it. Data integrity and safety in the execution rely on proof of existence.