【作者】 王谦；

【导师】 林文松；

【作者基本信息】 东南大学 ， 网络空间安全， 2022， 硕士

【摘要】 机器学习模型的构建依赖从多方收集训练数据,以实现对事件的检测、分类和预测。由于带宽、存储和隐私方面的考虑,将所有数据发送到一个集中的位置通常是不切实际的。联邦学习的客户端将本地训练的局部模型而不是原始数据上传到中心服务器,这增强了各个节点的隐私,并减少了传输大量数据带来的通信压力。然而,传统架构依赖于一个中心服务器的协调,存在信任危机和单点故障的风险。本研究基于区块链技术提出一种去中心化、防篡改的联邦学习架构,包括同步和异步两种训练模式。这种架构能够消除中心服务器的单点故障并降低外部攻击隐患,并兼容传统联邦学习架构的隐私保护方法。为了研究所提出模型的延迟消耗,本文分析了每轮全局聚合端到端的延迟情况,并在区块链的不同参数设置下测试了延迟结果。此外,针对如何利用有限的资源获得最佳学习性能的问题,本文提出了在有限时间下自适应调整本地训练和全局聚合次数的算法。实验结果表明,与固定迭代次数的参照组相比,本文提出的方法在动态环境中取得了更好的性能。更多还原

【Abstract】 Machine learning models are often built from the collected data,to enable the detection,classification,and prediction of future events.Due to bandwidth,storage,and privacy concerns,it is often impractical to send all the data to a centralized location.Clients involved in federated learning upload locally trained models to a central server instead of raw data,which enhances the privacy of individual nodes and reduces the communication stress caused by the transmission of large amounts of data.However,the traditional architecture relies on a central server with crises of trust and single point failure.In this thesis,we propose a decentralized and tamper-proof federated learning architecture based on blockchain technology,including both synchronous and asynchronous training modes.This architecture can eliminate the single point failure of the central server and reduce the potential for external attacks,and is compatible with the privacypreserving approach of traditional federated learning architectures.To investigate the latency consumption of the proposed model,an end-to-end latency model for each round of global aggregation is proposed and its latency results are tested under different parameter settings of the blockchain.In addition,we investigate how to achieve better model parameters with limited resources,and propose an algorithm that allocates adaptive adjustment of the number of local training and global aggregation under limited time.Experimental results show that our method is well adapted in dynamic environments compared to a reference group with a fixed number of iterations.更多还原