Internet of Things (IoT) interconnects a massive amount of devices, generating heterogeneous data with diverse characteristics. IoT data emerges as a vital asset for data-intensive IoT applications, such as healthcare, smart city and predictive maintenance, harnessing the vast volume of heterogeneous data to its maximum advantage. These applications leverage different Artificial Intelligence (AI) algorithms to discover new insights. While machine learning effectively uncovers implicit patterns through model training, centralizing IoT data for training poses significant privacy and security concerns. Federated Learning (FL) offers an promising solution, allowing IoT devices to conduct local learning without sharing raw data with third parties. Model-heterogeneous FL empowers clients to train models with varying complexities based on their hardware capabilities, aligning with heterogeneity of devices in real-world IoT environments. In this article, we review the state-of-the-art model-heterogeneous FL methods and provide insights into their merits and limitations. Moreover, we showcase their applicability to IoT and identify the open problems and future directions. To the best of our knowledge, this is the first article that focuses on the topic of model-heterogeneous FL for IoT.

翻译：物联网（IoT）实现了海量设备的互联，生成了具有多样特征的高度异构数据。物联网数据作为支撑医疗健康、智慧城市和预测性维护等数据密集型应用的重要资产，能够最大限度发挥异构海量数据的价值。这些应用利用不同的人工智能（AI）算法探索新见解。尽管机器学习通过模型训练能够有效发现隐含模式，但集中式训练物联网数据会引发严重的隐私与安全问题。联邦学习（FL）提供了一种有前景的解决方案，允许物联网设备在本地进行学习而无需向第三方共享原始数据。模型异构联邦学习使得客户端能够根据其硬件能力训练不同复杂度的模型，与真实物联网环境中设备的异构性高度契合。本文综述了当前最先进的模型异构FL方法，深入分析了其优势与局限性。此外，我们展示了这些方法在物联网中的适用性，并指出了开放性问题与未来研究方向。据我们所知，这是首个聚焦于面向物联网的模型异构FL研究主题的综述文章。