EvoIoT: An evolutionary IoT and non-IoT classification model in open environments

Abstract

IoT device identification is essential for both device asset management and security management. IoT device identification in open environments is the key to its application in real environments, where IoT and non-IoT device identification is the first and critical step. However, existing methods are either not applicable to open environments or have poor scalability and sustainability for IoT and non-IoT device identification. This paper presents EvoIoT, an IoT and non-IoT identification model designed to apply in open environments with high scalability and the ability to run sustainably and effectively. EvoIoT achieves high scalability by applying a unified model for all devices. To mine discriminative features from encrypted traffic, EvoIoT extracts original features from packet headers and judges feature importance in the entire model update process. For sustainability, EvoIoT proposes a representative device choosing method and model update method to address the concept drift caused by new types of devices. EvoIoT is the first high-performance model that systematically addresses the IoT and non-IoT device classification problem. We evaluate EvoIoT on two public datasets and a private dataset collected from a laboratory setting. The evaluation results show that EvoIoT is on average 6.27%∼48.89% more accurate than state-of-the-art methods.