IoT device identification is the process of recognizing and verifying connected IoT devices to the network. This is an essential process for ensuring that only authorized devices can access the network, and it is necessary for network management and maintenance. In recent years, machine learning models have been used widely for automating the process of identifying devices in the network. However, these models are vulnerable to adversarial attacks that can compromise their accuracy and effectiveness. To better secure device identification models, discretization techniques enable reduction in the sensitivity of machine learning models to adversarial attacks contributing to the stability and reliability of the model. On the other hand, Ensemble methods combine multiple heterogeneous models to reduce the impact of remaining noise or errors in the model. Therefore, in this paper, we integrate discretization techniques and ensemble methods and examine it on model robustness against adversarial attacks. In other words, we propose a discretization-based ensemble stacking technique to improve the security of our ML models. We evaluate the performance of different ML-based IoT device identification models against white box and black box attacks using a real-world dataset comprised of network traffic from 28 IoT devices. We demonstrate that the proposed method enables robustness to the models for IoT device identification.

翻译：物联网设备识别是识别并验证连接到网络的物联网设备的过程。这是确保只有授权设备能访问网络的关键环节，也是网络管理与维护的必要条件。近年来，机器学习模型被广泛用于网络设备识别任务的自动化。然而，这些模型易受对抗攻击的影响，可能导致其准确性与有效性受损。为更好保护设备识别模型，离散化技术能够降低机器学习模型对对抗攻击的敏感度，从而增强模型的稳定性与可靠性。另一方面，集成方法通过结合多个异构模型来减少模型中残留噪声或错误的影响。因此，本文融合了离散化技术与集成方法，并研究了其对模型抗对抗攻击鲁棒性的影响。换言之，我们提出一种基于离散化的集成堆叠技术，以提升机器学习模型的安全性。我们使用包含28个物联网设备网络流量的真实数据集，评估了不同基于机器学习的物联网设备识别模型在白盒与黑盒攻击下的性能。实验证明，所提方法能够增强物联网设备识别模型的鲁棒性。