Hierarchical enhanced deep encoder-decoder for intrusion detection and classification in cloud IoT networks

Abstract

Securing cloud-based internet of things (IoT) networks against intrusions and attacks is a significant challenge due to their complexity, scale, and the diverse nature of connected devices. IoT networks consist of billions of devices, computer servers, data transmission networks, and application computers, all communicating vast amounts of data that must adhere to various protocols. This study introduces a novel approach, termed hierarchical enhanced deep encoder-decoder with adaptive frequency decomposition (HED-EDFD), and is designed to address these challenges within cloud-based IoT environments. The HED-EDFD methodology integrates adaptive frequency decomposition, specifically adaptive frequency decomposition, with a deep encoder-decoder model. This integration allows for the extraction and utilization of frequency domain features from time-sequence IoT data. By decomposing data into multiresolution wavelet coefficients, the model captures both high-frequency transient changes and low-frequency trends, essential for detecting potential intrusions. The deep encoder-decoder model, enhanced with deep contextual attention mechanisms, processes these features to identify complex patterns indicative of malicious activities. The hierarchical structure of the approach includes a hierarchical wavelet-based attention mechanism, which enhances the accuracy and robustness of feature extraction and classification. To address the issue of imbalanced intrusion data, a cosine-based SoftMax classifier is employed, ensuring effective recognition of minority class samples.