Open Access

ARTICLE

An Improved Dream Optimization Algorithm-Driven Feature Selection Model for IoT Traffic Anomaly Detection

Hui Xu, Shuang Qu^*, Pan Hu

School of Computer Science and Artificial Intelligence, Hubei University of Technology, Wuhan, China

* Corresponding Author: Shuang Qu. Email:

(This article belongs to the Special Issue: Advances in IoT Security: Challenges, Solutions, and Future Applications, 2nd Edition)

Computers, Materials & Continua 2026, 89(2), 49 https://doi.org/10.32604/cmc.2026.087054

Received 09 June 2026; Accepted 28 July 2026; Issue published 15 September 2026

Abstract

With the rapid growth in the number of end devices in the Internet of Things (IoT), network traffic has become increasingly complex and redundant, while multiple attack types often coexist, posing major challenges to traffic anomaly detection. Traditional machine learning-based methods for IoT traffic anomaly detection often suffer from severe feature redundancy, high computational complexity, and low detection efficiency, making it difficult to simultaneously achieve high detection accuracy and computational efficiency. To address this issue, metaheuristic algorithms are often introduced in the feature selection stage to reduce feature redundancy and improve detection efficiency. However, the original Dream Optimization Algorithm (DOA) still has limitations when handling complex optimization problems such as feature selection for IoT traffic anomaly detection, including reduced population diversity in later iterations, insufficient global exploration capability, and a tendency to fall into local optima, which in turn affect the stability of feature selection and detection performance. Therefore, this paper proposes an Improved Dream Optimization Algorithm (IDOA) based on the original DOA, and further constructs an IDOA-based IoT traffic anomaly detection model for feature selection and classification tasks. Experimental results on benchmark functions and multiple publicly available datasets demonstrate the effectiveness and competitiveness of the proposed method. On four general-purpose classification benchmarks from the University of California, Irvine (UCI) Machine Learning Repository, IDOA demonstrates competitive feature-selection and classification performance. On the NF-ToN-IoT-v2 dataset, the binary classification accuracy reaches 98.01% with only 7.4 selected features on average, and the multiclass classification accuracy reaches 96.382% with 9.20 selected features on average. The experimental results indicate that the proposed method reduces feature redundancy and improves prediction efficiency in the evaluated offline setting while maintaining competitive detection accuracy.

---

Keywords

---