A Hierarchical Blockchain Architecture for Federated Learning in Edge Computing Networks

Abstract

Due to the growth of the Internet of Things (IoT) industry, numerous smart devices, such as wearable healthcare devices and smart home devices, are widely exploited. The adopted devices generate massive amounts of data, promoting high-quality machine learning (ML) technologies that can improve the performance of IoT. However, the generated data contains personal and sensitive information, and IoT devices are susceptible to attacks. Therefore, data privacy and security have become focused issues. Blockchained-federated learning (BFL) is one of the fastest-growing paradigms for building a trusted decentralized learning pattern for securing IoT systems in the edge computing environment. However, integrating blockchain and federated learning (FL) in edge computing networks faces some challenges. On the one hand, most BFL systems utilize the Proof of Work (PoW) consensus mechanism, consuming large amounts of computing power and adding extra latency. On the other hand, blockchain and FL are resource-intensive systems, adopting BFL systems efficiently in highly heterogeneous edge computing networks needs to be considered. In this dissertation, we design an efficient consensus protocol named Concordia for blockchain networks. The proposed protocol uses threshold signatures as a voting mechanism to confirm the validity of the proposed block. It requires only one round of one-way communication to achieve finality for each block. With a gossip-like communication scheme, the consensus can be reached within O(logN) steps. We also design a dual- mode consensus protocol named Flexico to improve the availability of the previous Concordia consensus protocol. Participants are divided into active and passive nodes. The consensus can be expedited through a fast-mode operation when the majority of the active nodes can communicate synchronously. Under non-ideal conditions, the backup protocol takes over the agreement process from its fast-mode counterpart without starting over the suspended round. The safety and liveness of the proposed protocol are guaranteed with lower communication costs, which balance the trade-off between protocol efficiency and availability. Then, we propose a novel multi-tiered blockchain-enabled semi- asynchronous FL architecture in edge computing networks. The blockchain consists of two layers, where the bottom layer is a Direct Acyclic Graph (DAG), and the top layer includes multiple sub-chains. The parallel DAG is adopted to record the local model uploaded by the training nodes for semi- asynchronous FL to solve the stragglers problem. Moreover, the uploaded local models are validated and aggregated by sub-chain nodes to prevent the system from model poisoning attack. By utilizing a two-layer blockchain architecture, the model training and validating are decoupled, enhancing the efficiency, security, and scalability of the system are improved.