Hybrid Deep Learning with Active Period Segmentation and Weighted Voting Ensemble for Intrusion Detection in IoT
Chitty Avula, Sathyanarayana Bachala · 2025
The Internet-of-Things (IoT) has revolutionized industries by enabling seamless communication, surveillance, and control, but its rapid growth has also brought significant security challenges such as data breaches, manipulation, and system failures. Existing intrusion detection systems (IDSs) often struggle with issues like data imbalance, insufficient contextual insights, and complexities in modeling sophisticated attack patterns. This paper introduces a novel IDS framework that overcomes these limitations through active period segmentation to remove inactive traffic noise, Word2Vec embedding for semantic feature generation, and statistical feature extraction to capture critical metrics. These refined features are processed by a hybrid deep learning architecture combining EfficientNet-NB7 and Bi-LSTM for advanced feature representation. For classification, a Weighted Voting Ensemble (WVE) integrating AdaBoost, Random Forest, and XGBoost adjusts classifier weights dynamically to enhance reliability and ensures that more reliable classifiers dominate the decision-making process, enhancing the ensemble’s overall predictive power. The proposed system achieves exceptional results with $99.94 \%$ accuracy, $99.72 \%$ precision, $99.77 \%$ recall, and a 99.29% F-Measure, demonstrating its robustness and effectiveness in safeguarding IoT infrastructures.