Υποστήριξη αποκεντρωμένης ομόσπονδης μάθησης μέσω Pos Blockchains

Abstract

In the case of Federated Learning (FL) there's a problem.Most existing systems require a central coordinator or permissioned ledgers, restricting the transparency of the data and leaving the prevention of Sybil attacks in a grey area.We have addressed these issues by using a permissionless Proof-of-Stake (PoS) blockchain to coordinate FL, and making it difficult for Sybil attacks to be carried out by putting model lists and updates directly onto the blockchain.Large amounts of data are instead stored offchain, using InterPlanetary File System (IPFS) which takes care of the problem of storage space.Our system has a training process that is split into rounds, with clients updating a shared model locally, sending out IPFS content identifiers to the network and a designated 'lister' pooling the updates and publishing the new global model.The idea is that the blockchain would be based on Proof of Stake with longest chain, highest stake finality, but our prototype mimics this with a tiny light-weight proof-of-stake mechanism.