Abstract
<title>Abstract</title> <p>This work presents a compact and computationally efficient intrusion detection system (IDS) pipeline for Internet of Things (IoT) environments, integrating hybrid data balancing, filter-based feature reduction, and meta-heuristic hyperparameter optimization to construct an efficient one-dimensional Convolutional Neural Network (1D CNN). Class imbalance is addressed using an SMOTE--ENN--LOF sequence. At the same time, feature redundancy is mitigated by the Bowerbird Courtship-Inspired Feature Selection (BBFS) algorithm, which ranks features using a composite score that combines clustering-based Mutual Information (CMI), Fisher Score, and a correlation penalty. The Hunger Games Search (HGS) algorithm is employed to derive compact 1D CNN architectures by jointly considering attack-detection performance, trainable parameter count, and multiply--accumulate (MAC) operations. The proposed BBFS--HGS--1D CNN pipeline is evaluated on four diverse IoT intrusion detection datasets---ACI-IoT\,2023, UQ-IoT\,2021, Edge-IIoTset, and WUSTL-IIoT. The final models used compact feature subsets of 20/76, 23/76, 18/76, and 10/38 features, respectively, with trainable parameter counts ranging from 63,937 to 105,281 and MAC estimates ranging from 188,224 to 364,992. While the aggregate accuracy, precision, recall, and F1-score show only marginal gains on some saturated datasets, the proposed pipeline improves the reliability of attack detection by reducing false negatives, with particular benefit for underrepresented attack categories. Comparative evaluation against six reimplemented IDS baselines shows that the proposed pipeline achieves competitive detection quality while using reduced feature subsets and compact CNN configurations, indicating its potential for future deployment-oriented IDS studies.</p>