Asynchronous federated learning of main-side chain collaboration for multi-level task scheduling in IoT

Currently, the Internet of Things (IoT) boasts a multitude of applications, and the count of devices is escalating rapidly. Therefore, IoT systems need smarter ways of collecting and processing data to provide decision-making services to the consumer more quickly. The combination of federated learning (FL) and blockchain offers a novel way to solve the value extraction of massive data in IoT. However, due to the limitation of device computational capacity, heterogeneity of computational resources, and the straggler effect caused by frequent state changes of nodes, those issues in present blockchain-based FL still exist, especially in the complex IoT environment where the problem is more obvious. Aiming at the resource-constrained IoT systems, we propose an asynchronous FL framework that leverages main-side chain collaboration within a dual-chain architecture and dynamically allocates clients into groups according to their processing capacity and network performance. Furthermore, this framework employs a multi-level task scheduling strategy based on the deadline-driven maximal completion time algorithm (DDMLTS) for FL, and a side-chain approach is integrated to enhance security assessment and node scheduling. Experiments show that the proposed framework reduces the average completion time by 10.9% and 8.8% compared with BAFL and PerFedS2, respectively, while achieving comparable model accuracy across the evaluated datasets.