[论文解读] A VHetNet-Enabled Asynchronous Federated Learning-Based Anomaly Detection Framework for Ubiquitous IoT

Weili Wang, Omid Abbasi+4 more|arXiv (Cornell University)|Mar 6, 2023

UAV Applications and Optimization被引用 4

一句话总结

本文提出了一种基于VHetNet的异步联邦学习（AFL）框架，利用高空平台站（HAPS）和无人机（UAVs）协同训练一个全局异常检测模型，以应对无处不在的物联网（IoT）系统中的挑战，采用差分隐私和基于CA2C的联合优化方法，实现设备关联、UAV选择与轨迹规划的协同优化。与基线方法相比，该方法在检测精度、能耗和训练时间方面均表现出优越性能。

ABSTRACT

Anomaly detection for the Internet of Things (IoT) is a major intelligent service required by many fields, including intrusion detection, device-activity analysis, and security supervision. However, the heterogeneous distribution of data and resource-constrained end nodes present challenges for existing anomaly detection models. Due to the advantages of flexible deployment and multi-dimensional resources, high altitude platform stations (HAPSs) and unmanned aerial vehicles (UAVs), which are important components of vertical heterogeneous networks (VHetNets), have significant potential for sensing, computing, storage, and communication applications in ubiquitous IoT systems. In this paper, we propose a novel VHetNet-enabled asynchronous federated learning (AFL) framework to enable decentralized UAVs to collaboratively train a global anomaly detection model. In the VHetNet-enabled AFL framework, a HAPS operates as a central aerial server, and the local models trained in UAVs are uploaded to the HAPS for global aggregation due to its wide coverage and strong storage and computation capabilities. We introduce a UAV selection strategy into the AFL framework to prevent UAVs with low local model quality and large energy consumption from affecting the learning efficiency and detection accuracy of the global model. To ensure the security of transmissions between UAVs and the HAPS, we add designed noise to local model parameters in UAVs to achieve differential privacy. Moreover, we propose a compound-action actor-critic (CA2C)-based joint device association, UAV selection, and UAV trajectory planning algorithm to further enhance the overall federated execution efficiency and detection model accuracy. Extensive experimental evaluation on a real-world dataset demonstrates that the proposed algorithm can achieve high detection accuracy with short federated execution time and low energy consumption.

研究动机与目标

• 解决无处不在的物联网系统在异常检测中面临的数据异构性与资源约束问题。
• 通过异步联邦学习（AFL）实现在UAV之间的去中心化、隐私保护式模型训练。
• 通过整合UAV轨迹规划、设备关联与UAV选择，构建统一的优化框架，提升学习效率与检测精度。
• 利用差分隐私的WGAN-GP，确保UAV与HAPS之间模型参数交换的安全性。
• 在有限机载能量条件下，最小化UAV辅助异常检测中的能耗与训练延迟。

提出的方法

• 采用VHetNet架构，以HAPS作为中心空中服务器，UAV作为移动边缘节点，负责本地模型训练。
• 采用异步联邦学习（AFL），允许每轮仅部分UAV参与，从而降低训练延迟与通信开销。
• 引入UAV选择策略，排除低质量或高能耗的UAV，提升全局模型收敛速度与精度。
• 通过在本地模型参数传输前添加校准噪声，实施差分隐私，确保数据隐私。
• 采用复合动作演员-评论家（CA2C）强化学习算法，在能量约束下联合优化设备关联、UAV选择与UAV轨迹规划。
• 使用差分隐私的WGAN-GP作为本地异常检测模型，以增强隐私保护与检测性能。

实验结果

研究问题

• RQ1 如何设计一种去中心化的异常检测框架，以应对无处不在的物联网系统中数据异构性与资源约束的挑战？
• RQ2 在UAV辅助的物联网系统中，异步联邦学习结合UAV选择对模型收敛性与检测精度有何影响？
• RQ3 如何有效将差分隐私集成到模型聚合过程中，以保护UAV-HAPS通信中的敏感数据？
• RQ4 设备关联、UAV选择与轨迹规划的联合优化在多大程度上能提升UAV赋能联邦学习中的能效与学习效率？
• RQ5 所提出的基于CA2C的调度策略在检测精度、能耗与训练时间方面相较于基线方法表现如何？

主要发现

• 所提出的CA2C-AFL算法在未见测试数据上实现了最高的检测精度，优于DDPG-FL、DQN-AFL与独立模型。
• CA2C-AFL算法的平均UAV能耗在所有对比算法中最低，归因于优化的轨迹规划与选择性模型更新。
• CA2C-AFL框架显著降低了平均与最大训练执行时间，得益于异步学习与高效的UAV子集选择，优于DDPG-FL与独立方法。
• 独立模型的检测精度在新数据集上显著下降，表明其泛化能力差，原因在于缺乏数据多样性与模型交换机制。
• DDPG-FL算法训练时间最长，因其要求所有UAV完成本地更新后才能进行全局聚合，违背了异步学习原则。
• 通过WGAN-GP实现的差分隐私集成，确保了模型参数传输的安全性，且未损害检测性能。