Using Large Language Models to Identify Devices on Your Network

What is it about?

Modern networks contain many internet-connected devices, such as smart speakers, cameras, TVs, and appliances. However, users and network administrators often do not know what devices are present on their networks or who made them. In this study, we develop an artificial intelligence system that identifies real-world Internet of Things (IoT) devices using the network information they naturally generate. We use data from thousands of households and train a language model to infer device manufacturers from network metadata, even when information is incomplete or inconsistent. Our system achieves high accuracy across more than 2,000 device vendors and remains effective when key identifying information is missing or misleading. We also evaluate the system on an independent dataset and show that it generalizes well to new environments. This work demonstrates how AI can improve network transparency, helping users and organizations better understand, manage, and secure the devices connected to their networks.

Featured Image

Photo by Aerps.com on Unsplash

Photo by Aerps.com on Unsplash

Why is it important?

As homes, workplaces, and shared spaces become increasingly connected, it is becoming harder to know what devices are present on a network and whether they can be trusted. Existing device identification systems often rely on incomplete information and struggle with the large variety of devices found in real-world environments. This work demonstrates that modern AI models can accurately identify thousands of device manufacturers using only network metadata, even when information is missing or inconsistent. Unlike traditional approaches, our method can reason across multiple sources of evidence and generalize to rare or previously unseen devices. By improving visibility into connected devices, this research can support network security, privacy protection, asset management, and the detection of unauthorized or suspicious devices in homes and organizations.