Hybrid Tip Selection for DAG-based Federated Learning over Mobile Network

Abstract

Federated learning (FL) enables clients to participate in machine learning tasks in a private way. Applying blockchain into FL for decentralization and security has attracted much attention recently. The blockchain with a directed acyclic graph (DAG) structure enables mobile devices to participate in decentralized FL more flexibly while reducing resource consumption and is more suitable for implementing decentralized FL in mobile networks than traditional blockchains. Non-independent and identically distributed (non-IID) data is a common problem in FL. Existing work on DAG-based FL lacks a suitable optimization method for non-IID data. In this paper, we briefly describe a DAG-based FL approach in mobile networks. In order to mitigate the negative effects of non-IID data, consensus in DAG is improved by utilizing a new tip (Unconfirmed blocks in the DAG ledger) selection algorithm proposed in this paper to help clients find suitable models more easily in DAG-based FL. Experiments on multiple datasets show that the method proposed in this paper has better results than existing work and is closer to traditional FL.