Sharma Anshika, Rani Shalli, Shabaz Mohammad
Chitkara University Institute of Engineering and Technology, Chitkara University, Rajpura, Punjab, India.
Marwadi University Research Center, Department of Computer Engineering, Faculty of Engineering and Technology, Marwadi University, Rajkot, Gujarat, India.
PLoS One. 2025 Aug 1;20(8):e0329227. doi: 10.1371/journal.pone.0329227. eCollection 2025.
With the expansion of Internet of Things (IoT) devices, security is an important issue as attacks are constantly gaining more complex. Traditional attack detection methods in IoT systems have difficulty being able to process real-time and access limitations. To address these challenges, a stacking-based Tiny Machine Learning (TinyML) models has been proposed for attack detection in IoT networks. This ensures detection efficiently and without additional computational overhead. The experiments have been conducted using the publicly available ToN-IoT dataset, comprising a total of 461,008 labeled instances with 10 types of attacks categories. Some amount of data preprocessing has been done applying methods such as label encoding, feature selection, and data standardization. A stacking ensemble learning technique uses multiple models combining lightweight Decision Tree (DT) and small Neural Network (NN) to aggregate power of the system and generalize. The performance of the model is evaluated by accuracy, precision, recall, F1-score, specificity, and false positive rate (FPR). Experimental results demonstrate that the stacked TinyML model is superior to traditional ML methods in terms of efficiency and detection performance, and its accuracy rate is 99.98%. It has an average inference latency of 0.12 ms and an estimated power consumption of 0.01 mW.
随着物联网（IoT）设备的扩展，安全成为一个重要问题，因为攻击手段日益复杂。物联网系统中的传统攻击检测方法难以处理实时性和访问限制问题。为应对这些挑战，提出了一种基于堆叠的 Tiny 机器学习（TinyML）模型用于物联网网络中的攻击检测。这确保了高效检测且无额外计算开销。实验使用公开可用的 ToN-IoT 数据集进行，该数据集共有 461,008 个带标签的实例，包含 10 种攻击类别。应用标签编码、特征选择和数据标准化等方法进行了一定量的数据预处理。一种堆叠集成学习技术使用多个模型，结合轻量级决策树（DT）和小型神经网络（NN）来聚合系统能力并进行泛化。通过准确率、精确率、召回率、F1 值、特异性和误报率（FPR）来评估模型性能。实验结果表明，堆叠的 TinyML 模型在效率和检测性能方面优于传统机器学习方法，其准确率为 99.98%。它的平均推理延迟为 0.12 毫秒，估计功耗为 0.01 毫瓦。
