Using Siamese Neural Networks for Efficient and Accurate IoT Device Identification

With the wide adoption of Internet of Things (IoT) devices, it becomes crucial to identify which IoT devices are connected to the network at a specific time. Previous studies have built machine learning models that can accurately identify IoT devices on a specific network based on their traffic characteristics. However, one limitation of such models is that whenever a new device joins the network, the model has to be retrained from scratch, which adds a lot of computation overhead. In this work, we propose the use of Siamese Neural Networks to reduce the retraining frequency of IoT device identification models. We use a public dataset containing traffic features from 10 devices. To validate the proposed idea, we first compare the performance of classical multi-class classification neural networks with Siamese Networks on the task at hand. We see that both networks perform similarly. Then, we build 10 separate models based on Siamese networks, and we train each of them to recognize a different combination of 9 devices. Then, we use each of the trained models to recognize the device that was not part of the training set. We assess the performance of each model, and we compare the results with the ones achieved by the multi-class classification network. We prove that with the proposed approach, similar or even better outcomes are achieved, with the main advantage of not having to retrain. Finally, we test the proposed approach against 2 other datasets: Aalto and UNSW. We compare the outcomes with previous works, and we prove that Siamese Networks achieve a better performance.