10 March 2025 in IEEE Data Descriptions
This work presents the Deakin IoT Traffic dataset comprising of packets emitted by Internet of Things (IoT) devices, each having different functions and behaviors. In total, 112 million packets were collected over 119 days from 24 IoT devices. Of the 24 IoT devices, 19 are distinct. The collected packets are stored in PCAP (Packet Capture) files, preserving a wide array of network protocols such as DHCP, ARP, DNS, HTTP, and other TCP/UDP-based applications, which enable a detailed analysis of communication patterns. Alongside the IoT PCAP files, we provide the full PCAP files which contain traffic from 36 non-IoT devices. In addition to the PCAP files, the dataset includes a CSV file that maps each IoT device to its unique MAC address, simplifying device-specific analysis. To assist researchers, Python scripts are provided for extracting and processing packets, demonstrating ways to leverage the data for various applications. This dataset is valuable for a wide range of research purposes: it enables researchers to study and differentiate network behaviors based on device function, supports behavior-based profiling for anomaly detection, and provides a foundation for designing IoT-tailored network policies.
16 December 2024 in IEEE Transactions on Machine Learning in Communications and Networking
Machine Learning (ML)-based techniques are increasingly used for network management tasks, such as intrusion detection, application identification, or asset management. Recent studies show that neural network-based traffic analysis can achieve performance comparable to human feature-engineered ML pipelines. However, neural networks provide this performance at a higher computational cost and complexity, due to high-throughput traffic conditions necessitating specialized hardware for real-time operations. This paper presents lightweight models for encoding characteristics of Internet-of-Things (IoT) network packets; 1) we present two strategies to encode packets (regardless of their size, encryption, and protocol) to integer vectors: a shallow lightweight neural network and compression. With a public dataset containing about 8 million packets emitted by 22 IoT device types, we show the encoded packets can form complete (up to 80%) and homogeneous (up to 89%) clusters; 2) we demonstrate the efficacy of our generated encodings in the downstream classification task and quantify their computing costs. We train three multi-class models to predict the IoT class given network packets and show our models can achieve the same levels of accuracy (94%) as deep neural network embeddings but with computing costs up to 10 times lower; 3) we examine how the amount of packet data (headers and payload) can affect the prediction quality. We demonstrate how the choice of Internet Protocol (IP) payloads strikes a balance between prediction accuracy (99%) and cost. Along with the cost-efficacy of models, this capability can result in rapid and accurate predictions, meeting the requirements of network operators.
1 October 2026 in ACM Transactions on Internet of Things
Systematically managing the cyber health of Internet of Things (IoT) networks at scale requires understanding the functions and behaviors of connected devices. Current IoT device management approaches rely on classifying these behaviors based on the make-and-model of devices. However, these make-and-model classes proliferate over time and do not generalize well to unseen devices or networks. To overcome this, we propose a method to classify IoT behaviors into higher-abstraction metaclasses. Our contributions are three-fold: (1) We develop a method based on Large Language Models (LLMs) to derive metaclasses, grouping 125 IoT make-and-model types into seven functionality-based classes, using public datasets collected from five IoT networks. (2) We train PacketTransformer—a GPT-2 classifier—on data from one network and evaluate its generalizability on four unseen networks. In our cross-network setting, classification using metaclasses achieves up to 38% higher accuracy than make-and-model classification. (3) We investigate performance-improvement strategies beyond the multiclass baseline by training separate models for each metaclass and by calibrating stacked one-class outputs. These experiments improve performance in several settings, with the strongest one-class results achieving balanced accuracies of 71% for Cameras, 75% for Speakers, and 79% for Sensors.