跳转至

异构边缘-云连续统中的自适应人工智能任务划分与安全卸载

文章背景与核心概要

随着物联网设备的普及，在资源受限的边缘端运行人工智能模型的应用需求日益增长。然而，传统的AI任务划分与卸载方法大多依赖于静态策略，无法适应运行时动态变化的网络环境，且缺乏在真实硬件上的实测验证。为此，本文提出了一种全新的自适应框架，能够在异构边缘-云连续统中动态拆分和卸载神经网络层。该框架在启动时对模型进行性能剖析，实时监测网络链路状况，并定期更新划分策略，以应对环境波动。

在包含树莓派边缘设备、笔记本雾端节点以及高性能桌面云的物理测试床进行的实验表明，该方法在运行 VGG16、AlexNet 和 MobileNetV2 等经典卷积神经网络时表现优异。相比静态基准方法，该框架实现了 27.09% 至 35.82% 的能耗降低 以及 6.34% 至 22.92% 的端到端延迟降低，充分证明了自适应划分相比传统静态方法的显著优势。

摘要 (Summary)

本文介绍了一种自适应框架，用于在异构边缘-云连续统（从资源受限的树莓派等物联网设备到高性能桌面云）中拆分和卸载神经网络层。与依赖模拟的传统静态方法不同，该框架在启动时对模型进行性能剖析，监测运行时的网络状况，并定期更新划分以应对环境波动。通过在真实硬件上使用 VGG16、AlexNet 和 MobileNetV2 进行测试，该方法相比静态基准取得了显著改进：能耗降低 27.09%–35.82%，端到端延迟降低 6.34%–22.92%。

This paper introduces an adaptive framework for splitting and offloading neural network layers across a heterogeneous edge-cloud continuum (ranging from resource-constrained IoT devices like Raspberry Pis to high-performance desktop clouds). Unlike traditional static methods that rely on simulations, this framework profiles models at startup, monitors runtime network conditions, and periodically updates partitions to handle environmental fluctuations. Tested on real hardware using VGG16, AlexNet, and MobileNetV2, the approach yields significant improvements over static baselines: a 27.09%–35.82% reduction in energy consumption and a 6.34%–22.92% reduction in end-to-end latency.

文章详情 (Article Details)

  • arXiv ID: arXiv:2605.09623 [cs.DC]
  • 主要主题: 分布式、并行与集群计算 (cs.DC)
  • 其他主题: 人工智能 (cs.AI)、机器学习 (cs.LG)、网络与互联网架构 (cs.NI)、性能 (cs.PF)
  • 提交日期: 2026年5月10日提交；2026年8月18日最后修订 (v2)
  • 相关 DOI: 10.1007/978-3-032-35576-8_23
  • arXiv ID: arXiv:2605.09623 [cs.DC]
  • Primary Subject: Distributed, Parallel, and Cluster Computing (cs.DC)
  • Other Subjects: Artificial Intelligence (cs.AI), Machine Learning (cs.LG), Networking and Internet Architecture (cs.NI), Performance (cs.PF)
  • Submission Dates: Submitted on 10 May 2026; Last revised 18 Aug 2026 (v2)
  • Related DOI: 10.1007/978-3-032-35576-8_23

作者 (Authors)

  • Akuen Akoi Deng
  • Eimantas Butkus
  • Alfreds Lapkovskis
  • Praveen Kumar Donta
  • Akuen Akoi Deng
  • Eimantas Butkus
  • Alfreds Lapkovskis
  • Praveen Kumar Donta

摘要正文 (Abstract)

近年来，在资源受限的物联网设备上使用人工智能的应用规模显著增长。然而，现有的跨边缘-云连续统的 AI 任务划分与卸载方法通常依赖于忽略运行时动态的静态方法。此外，它们往往在模拟环境中进行评估，而非真实的硬件。为了弥补这一空白，我们提出了一种在异构连续统中动态拆分神经网络层的框架。该框架在启动时对模型进行剖析，测量节点之间的网络链路状况，并定期重新评估划分以适应环境变化。我们创建了一个物理测试床，包含一个树莓派边缘设备、一台笔记本雾端设备以及一台作为云的高性能台式电脑。我们对三个广受欢迎的卷积神经网络（VGG16、AlexNet 和 MobileNetV2）评估了该框架。我们的结果表明，与静态划分基准相比，该框架实现了能源和端到端延迟分别降低 27.09–35.82% 和 6.34–22.92%。这些发现证实了自适应划分优于静态划分。

In recent years, the use of artificial intelligence on resource-constrained IoT devices has grown significantly. However, existing approaches to AI task partitioning and offloading across the edge-cloud continuum typically rely on static methods that ignore runtime dynamics. Furthermore, they are often evaluated in simulated environments rather than on real hardware. To address this gap, we propose a framework that dynamically splits neural network layers across the heterogeneous continuum. The framework profiles the model at startup, measures network link conditions between nodes, and periodically re-evaluates the partition to adapt to environmental changes. We created a physical testbed comprising a Raspberry Pi edge device, a laptop fog, and a high-performance desktop PC as the cloud. We evaluated the framework over three widely adopted convolutional neural networks: VGG16, AlexNet, and MobileNetV2. Our results show that the framework achieves reductions in energy and end-to-end latency of 27.09--35.82% and 6.34--22.92%, respectively, compared to a static partitioning baseline. These findings confirm the superiority of adaptive to static partitioning.

访问与资源 (Access & Resources)

外部参考与引用 (External References & Citations)