Active Period Segmentation Assisted Semantically Embedded Hybrid Deep Feature Learning Environment for Intrusion Detection in IOT

Abstract

The rapid evolution of IoT technologies has enabled seamless, real-time communication and control across critical sectors like defense, smart cities, and healthcare. However, the growing scale of applications and users has introduced significant security risks, with intrusions threatening both data integrity and system reliability. Existing IDS solutions based on machine and deep learning often struggle with issues such as class imbalance, gradient instability, and limited contextual understanding. To address these challenges, this paper introduces a robust IDS model that combines active period segmentation with a semantically enhanced hybrid deep learning framework. The model efficiently eliminates inactive data, employs Word2Vec for semantic embedding, and extracts meaningful statistical features to enhance intrusion detection performance. These features are further refined through cross-correlation and processed using EfficientNet-NB7 and Bi-LSTM networks for deep feature extraction. Final classification is performed using an ensemble-of-ensembles (E2E) model that combines AdaBoost, Random Forest, and XGBoost, leveraging majority voting for decision making. The proposed IDS attain 99.91% accuracy, 99.62% precision, 99.70% recall, and a 99.23% F-measure, surpassing existing methods and providing robust security for IoT environments.