A DAG-Blockchain-Assisted Federated Learning Framework in Wireless Networks: Learning Performance and Throughput Optimization Schemes

In this article, an efficient wireless federated learning (FL) framework based on blockchain (BC) assistance is studied. Many existing frameworks adopt lots of third-party servers as consensus nodes, which is vulnerable to collusion attacks. In our framework, the blockchain-assisted FL (BFL) model selects edge users as blockchain nodes without any third-party intervention. Besides, the convergence analysis of this FL algorithm considering transmission outages is provided to prove the effects of wireless factors on FL. To solve the low efficiency of the BC based on the conventional linear chain structure, the Directed Acyclic Graph (DAG) blockchain is introduced into our work. Moreover, since the design and optimization of FL and BC in most existing works are done separately, this may result in sub-optimal performance. To achieve an excellent trade-off between FL efficiency and BC performance, a joint optimization problem regarding DAG-BFL is formulated. The optimization objective is to maximize the FL performance and DAG-BC throughput. To solve the complex nonconvex optimization problem, considering the resource-constrained BFL system, the joint communication and computing resource allocation as well as block designing schemes are proposed, which are based on the twin-loop penalty dual decomposition (PDD) method and the successive block minimization technique (BSUM). Extensive simulations are performed to demonstrate the effectiveness of the proposed method. Particularly, compared with the traditional alternative optimization, the proposed PDD-based algorithm achieves better performance.

I. Introduction

With the explosive growth of user data in wireless networks and the vigorous development of machine learning (ML) and artificial intelligence technologies [1], various cloud-based intelligent services are widely used in wireless networks [2]. Although traditional machine learning methods can provide intelligent services and bring great convenience to humans, they also pose more privacy leakage risks for us [3]. For example, the exposure of users' personal information, medical privacy information, and sensitive data on social media has caused considerable worry to us. Federated learning (FL) provides a decentralized approach that only trains ML models on various local users without uploading their personal data to a data processing center [4]. FL can effectively protect data privacy while providing intelligent services for users, and has received widespread attention from the business and academic communities.