Scholarlink

<p>Recently, the industrial and residential sectors have seen a rise in the popularity of internet of things (IoTs)-based applications. Most IoT devices are susceptible to various intrusions and have limited computation and storage capabilities due to their portable nature. Consequently, it is mandatory to have effective anomaly detection (AD) techniques to discern between real and fraudulent IoT data. Numerous research studies have been carried out using machine learning (ML) algorithms to develop IoT detection frameworks to find abnormal sensor data. However, high false-positive rates (FPRs) and computational complexity limits the performance. Intending to develop a novel AD for IoT, this paper suggests a lightweight TriFusion-AnomNet. It begins with collecting data from various sources and integrating it into a centralized repository. Using imputation and smoothing, a pre-processing phase is added to handle the missing values. Feature extraction is carried out to refine significant features via network-based and statistical features, and temporal features via recurrent neural network (RNN), and then normalize the data via Z-score. Besides, feature selection is applied to reduce dimensionality through proposed grid search-aided Bayesian optimization (GSBO). An ensemble-based approach TriFusion-AnomNet is proposed by fusing predictions from isolation forest (iForest), autoencoders (AE), and one-class support vector machine (OCSVM) to improve robustness. Finally, a feedback loop to refine and improve AD models based on human reviews and system performance. The suggested model attained 99.51% accuracy and outperformed state-of-the-art (SOTA) models.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Dynamic anomaly detection in IoT environments: a multi-model machine learning approach

  • 期刊论文

  • 发表日期：

    2026年7月3日

查看全文
  • Mohammad Tabrez Quasim

摘要

Recently, the industrial and residential sectors have seen a rise in the popularity of internet of things (IoTs)-based applications. Most IoT devices are susceptible to various intrusions and have limited computation and storage capabilities due to their portable nature. Consequently, it is mandatory to have effective anomaly detection (AD) techniques to discern between real and fraudulent IoT data. Numerous research studies have been carried out using machine learning (ML) algorithms to develop IoT detection frameworks to find abnormal sensor data. However, high false-positive rates (FPRs) and computational complexity limits the performance. Intending to develop a novel AD for IoT, this paper suggests a lightweight TriFusion-AnomNet. It begins with collecting data from various sources and integrating it into a centralized repository. Using imputation and smoothing, a pre-processing phase is added to handle the missing values. Feature extraction is carried out to refine significant features via network-based and statistical features, and temporal features via recurrent neural network (RNN), and then normalize the data via Z-score. Besides, feature selection is applied to reduce dimensionality through proposed grid search-aided Bayesian optimization (GSBO). An ensemble-based approach TriFusion-AnomNet is proposed by fusing predictions from isolation forest (iForest), autoencoders (AE), and one-class support vector machine (OCSVM) to improve robustness. Finally, a feedback loop to refine and improve AD models based on human reviews and system performance. The suggested model attained 99.51% accuracy and outperformed state-of-the-art (SOTA) models.