Artificial Intelligence (AI) based models are increasingly deployed in the Internet of Things (IoT), paving the evolution of the IoT into the AI of things (AIoT). Currently, the predominant approach for AI model training is cloud-centric and involves the sharing of data with external parties. To preserve privacy while enabling collaborative model training across distributed IoT devices, the machine learning paradigm called Federated Learning (FL) has been proposed. The future FL network is envisioned to involve up to millions of distributed IoT devices involved in collaborative learning. However, communication failures and dropouts by nodes can lead to inefficient FL. Inspired by the UAV-assisted communications in 5G heterogeneous networks (HetNet), we propose the UAV-assisted FL in this article. The FL model owner may employ UAVs to provide the intermediate model aggregation in the sky and mobile relay of the updated model parameters from data owners to the model owner. This therefore increases the reach of FL to data owners that face uncertain network conditions and improves the communication efficiency. To incentivize the UAV service providers, we adopt the multi-dimensional contract incentive design as a case study. The incentive compatibility of the contract ensures that the UAVs only choose an incentive package corresponding to its type, for example, traveling cost. The simulation results show that the UAV-assisted FL achieves significant improvement in communication efficiency and validates the incentive compatibility of our contract design.

中文翻译：

---

物联网时代无人机辅助通信高效联邦学习

基于人工智能 (AI) 的模型越来越多地部署在物联网 (IoT) 中，为 IoT 向物联网 (AIoT) 的演进铺平了道路。目前，人工智能模型训练的主要方法是以云为中心的，涉及与外部各方共享数据。为了在跨分布式物联网设备进行协作模型训练的同时保护隐私，已经提出了称为联合学习 (FL) 的机器学习范式。预计未来的 FL 网络将涉及多达数百万个参与协作学习的分布式物联网设备。然而，节点的通信失败和丢失会导致 FL 效率低下。受 5G 异构网络（HetNet）中无人机辅助通信的启发，我们在本文中提出了无人机辅助 FL。FL 模型所有者可以使用 UAV 来提供空中中间模型聚合和更新模型参数从数据所有者到模型所有者的移动中继。因此，这增加了 FL 对面临不确定网络条件的数据所有者的影响，并提高了通信效率。为了激励无人机服务提供商，我们采用多维合约激励设计作为案例研究。合约的激励兼容性确保无人机只选择与其类型相对应的激励包，例如旅行成本。仿真结果表明，无人机辅助 FL 实现了通信效率的显着提高，并验证了我们合约设计的激励兼容性。