【Author】
Yuan, Shuo; Cao, Bin; Peng, Mugen; Sun, Yaohua
【Source】2021 IEEE WIRELESS COMMUNICATIONS AND NETWORKING CONFERENCE (WCNC)
【Abstract】Despite the advantages of Federated Learning (FL), such as devolving model training to intelligent devices and preserving data privacy, FL still faces the risk of the single point of failure and attack from malicious participants. Recently, blockchain is considered a promising solution that can transform FL training into a decentralized manner and improve security during training. However, traditional consensus mechanisms and architecture for blockchain can hardly handle the large-scale FL task due to the huge resource consumption, limited throughput, and high communication complexity. To this end, this paper proposes a two-layer blockchain-driven FL framework, called as ChainsFL, which is composed of multiple Raft-based shard networks (layer-1) and a Direct Acyclic Graph (DAG)-based main chain (layer-2) where layer-1 limits the scale of each shard for a small range of information exchange, and layer-2 allows each shard to update and share the model in parallel and asynchronously. Furthermore, FL procedure in a blockchain manner is designed, and the refined DAG consensus mechanism to mitigate the effect of stale models is proposed. In order to provide a proof-of-concept implementation and evaluation, the shard blockchain base on Hyperledger Fabric is deployed on the self-made gateway as layer-1, and the self-developed DAG-based main chain is deployed on the personal computer as layer-2. The experimental results show that ChainsFL provides acceptable and sometimes better training efficiency and stronger robustness comparing with the typical existing FL systems.
【标题】ChainsFL：区块链驱动的从设计到实现的联邦学习
【摘要】尽管联邦学习 (FL) 具有将模型训练转移到智能设备和保护数据隐私等优势，但 FL 仍然面临单点故障和恶意参与者攻击的风险。最近，区块链被认为是一种很有前途的解决方案，可以将 FL 训练转变为去中心化的方式，并提高训练期间的安全性。然而，传统的区块链共识机制和架构由于资源消耗巨大、吞吐量有限、通信复杂度高等问题，难以处理大规模的 FL 任务。为此，本文提出了一个由多个基于 Raft 的分片网络（layer-1）和一个基于直接无环图（DAG）的主链（layer -2) 其中第 1 层限制每个分片的规模以进行小范围的信息交换，第 2 层允许每个分片并行和异步地更新和共享模型。此外，设计了区块链方式的FL程序，并提出了改进的DAG共识机制以减轻陈旧模型的影响。为了提供概念验证的实现和评估，基于Hyperledger Fabric的分片区块链部署在自制网关上作为layer-1，并在个人计算机上部署基于DAG的自研主链作为第 2 层。实验结果表明，与典型的现有 FL 系统相比，ChainsFL 提供了可接受的、有时甚至更好的训练效率和更强的鲁棒性。
评论