Assessing the Effectiveness of Siamese Neural Networks to Mitigate Frequent Retraining in IoT Device Identification Models

The rapid proliferation of the Internet of Things (IoT) requires the development of effective methods to identify devices over a network. Although traditional machine learning and deep learning approaches have been successful in this task, they require continuous retraining when new devices join the IoT network. To address this challenge, Siamese Neural Networks (SNNs) have been proposed, demonstrating strong performance without the need for model retraining when a new device is added to the network. In this work, we further assess the reliability of SNNs in reducing the necessity for model retraining when multiple devices join the network, as well as their generalization capabilities. Our findings indicate that while performance slightly declines with an increasing number of newly added devices, it remains satisfactory overall. Furthermore, we demonstrate the generalization of SNNs by training the model on one dataset and evaluating its performance when incorporating devices from a different dataset without retraining. The results highlight the robustness of SNNs and their ability to recognize devices on a different IoT network than the one they were initially trained on, and most importantly, without retraining. Finally, we show how the quality of the initial training set can affect the generalization capabilities of SNNs.