Scatter Protocol: an Incentivized and Trustless Protocol for Decentralized Federated Learning

Federated Learning is a form of privacy-preserving machine learning where multiple entities train local models, which are then aggregated into a global model. Current forms of federated learning rely on a centralized server to orchestrate the process, leading to issues such as requiring trust in the orchestrator, the necessity of a middleman, and a single point of failure. Blockchains provide a way to record information on a transparent, distributed ledger that is accessible and verifiable by any entity. We leverage these properties of blockchains to produce a decentralized, federated learning marketplace-style protocol for training models collaboratively. Our core contributions are as follows: first, we introduce novel staking, incentivization, and penalization mechanisms to deter malicious nodes and encourage benign behavior. Second, we introduce a dual-layered validation mechanism to ensure the authenticity of the models trained. Third, we test different components of our system to verify sufficient incentivization, penalization, and resistance to malicious attacks.