Personalized federated learning via directed acyclic graph based blockchain

Abstract

Abstract Common federated learning (FL) lacks consideration of clients' personalized requirements, which performs poorly for the scenario with data and resource heterogeneity. In order to overcome the challenge of heterogeneous characteristics, this letter proposes a novel decentralized personalized federated learning (PFL) architecture that first utilizes a directed acyclic graph (DAG) blockchain technology to achieve PFL efficiently, which is called PFLDAG. Simulation results demonstrate that PFLDAG approximately improves accuracy by 80% compared with the classic Google FedAvg algorithm, and by 10% compared with IFCA cluster PFL which considers personalized requirements. In addition, the approach also substantially improves the convergence speed.