【作者】 袁鑫；

【导师】 张彦；

【作者基本信息】 电子科技大学 ， 信息与通信工程， 2022， 硕士

【摘要】 在全球数据爆炸式增长的驱动下,人工智能在计算机视觉,自然语言处理以及无线通信网络等领域得到广泛应用。为了实现高效获取并利用大量移动终端所产生的数据,边缘智能框架应运而生,旨在将传统的集中式的云端智能迁移到距离移动用户更近的基站侧,但如何保护用户的数据隐私是边缘智能中一个很大的挑战。而联邦学习作为一种分布式的机器学习方法,通过服务器与节点间的参数通信来完成模型的训练,在充分利终端设备数据的同时能够保护用户数据隐私。然而,在移动边缘计算场景下,由于节点之间的通信资源和计算资源是异构的,若采用传统的联邦学习机制会导致模型学习的效率受限于速度最慢的节点。另外,联邦学习过程中部分节点因资源受限或隐私保护等原因,会上传旧的或错误的模型参数至服务器,同样会影响模型的训练效率。因此边缘服务器在联邦学习的过程需要进行节点选择,然而节点的资源异构性和节点间的数据异构性都会对节点选择造成影响。针对以上问题,本文在现有的联邦学习算法的基础上,提出了边缘智能网络中的联邦学习节点选择机制,本文的主要贡献如下:(1)针对节点资源异构和数据异构性,本文提出了基于沙普利值方法的节点贡献度评估策略,将节点在不同维度资源上的差异性用统一的贡献度衡量。针对节点上传旧的或错误的模型参数至服务器问题,本文提出了基于贡献度评估策略的恶意节点检测机制,并在数值仿真阶段验证了该机制的有效性。(2)针对传统联邦学习在移动边缘计算场景下学习效率低下问题,基于节点贡献度评估策略和恶意节点检测机制,本文提出了基于空间维度的节点选择和资源分配优化问题,提高模型学习效率的同时降低系统的能耗,并针对不同的场景给出求解算法,仿真结果表明在非独立同分布数据下本文所提的基于空间维度的节点选择方案在性能表现上相比于传统的方案提升6.4%。(3)为进一步优化联邦学习在移动边缘计算场景下的模型收敛时间,本文提出了基于时间维度的节点选择和资源分配方案,仿真结果表明基于时间维度的方案在模型收敛时间上相比于其他的方案缩短32.8%。(4)将上述两种节点选择方案结合,本文提出基于时空的节点选择和资源分配优化问题,仿真结果表明基于时空的节点选择方案相比与其他两种方案在模型的性能表现上提升了1.32%。更多还原

【Abstract】 Driven by the explosive growth of global data,artificial intelligence is widely used in fields such as computer vision,natural language processing,and wireless communication networks.In order to efficiently and quickly acquire and utilize the data generated by a large number of mobile device,the edge intelligence framework is proposed,the traditional centralized cloud intelligence is migrated to the base station side closer to the mobile user,but how to protect the user’s data privacy is a big challenge in edge intelligence.Federated learning enables distributed agents to train a common and shared model,making full use of terminal device data while protecting user data privacy.However,in the mobile edge computing scenario,since the communication resources and computing resources between nodes are heterogeneous,the efficiency of model learning depends on the least efficient node in traditional federated learning.In addition,during the federated learning process,some nodes may upload old or wrong model parameters to the server due to resource constraints or privacy protection,which greatly affects the training efficiency of the model.Therefore,the edge server needs to select nodes in the process of federated learning.However,the resource heterogeneity of nodes and the data heterogeneity between nodes affects node selection.Aiming at the above problems,based on the existing federated learning algorithms,this paper proposes a federated learning node selection mechanism in edge intelligent networks.The main contributions of this paper are as follows:(1)Aiming at the heterogeneity of node resources and data,this paper proposes a node contribution evaluation strategy based on the Shapley Value method,which measures the differences of nodes in resources of different dimensions with a unified contribution.Aiming at the problem of nodes uploading old or wrong model parameters to the server,this paper proposes a malicious node detection mechanism based on the contribution evaluation strategy,and verifies the effectiveness of the mechanism in the numerical simulation stage.(2)Aiming at the low learning efficiency of traditional federated learning in the mobile edge computing scenario,based on the node contribution evaluation strategy and malicious node detection mechanism,this paper proposes a spatial dimension-based node selection and resource allocation optimization problem to improve the model learning efficiency while reducing the system energy consumption.Compared with the traditional scheme,the performance of the node selection scheme based on spatial dimension is improved by 6.4% under non-IID data.(3)In order to further optimize the convergence time of federated learning in the mobile edge computing scenario,this paper proposes a node selection and resource allocation scheme based on the time dimension.The simulation results show that the time dimension-based scheme can shorten the model convergence time by 32.8% compared with other schemes.(4)Combining the above two node selection schemes,a space-time-based node selection and resource allocation optimization problem is proposed.The simulation results show that the performance of the space-time-based node selection scheme is improved by1.32% compared with the other two schemes.更多还原