Multidimensional Feature Learning Enhancement in IoT Intrusion Detection: An Adaptive Cost-Sensitive Autoencoder and Weighted Ensemble Approach

Internet of Things (IoT) defense against malicious activities heavily depends on intrusion detection systems (IDS). But unbalanced datasets and fast changing network conditions, prevalent in IoT applications, are common challenges for traditional IDS approaches. To address this challenge, in this paper we introduce a novel IDS framework for real-time IoT applications using a cost-sensitive autoencoder (CSAE) based ensemble method to enhance feature learning and IDS performance. In addition to the original data, we employ multidimensional data representation from various autoencoder layers to train machine learning (ML) models (e.g., Decision Trees (DT) and XGBoost (XGB)) which are further augmented by ensemble techniques. Moreover, we propose a strategic pretraining of the autoencoder using a different dataset, which promotes efficient information transfer and system flexibility. Empirical evaluations using two factual IoT datasets demonstrate a significant performance improvement over state-of-the-art IoT intrusion detection systems. Our IDS attains an impressive 99.99% accuracy, in multi-class configurations and a flawless 100% in binary classifications for the CICIoT2023 dataset. Similarly, for the DS2OS dataset, after finetuning the pretrained network, accuracy measures 99.88% in multi-class scenarios and 99.96% in binary classifications. The proposed cost sensitive autoencoder based weighted average (CSAE-WA) ensemble model outperformed six state-of-the-art intrusion detection techniques in a rigorous comparative analysis, demonstrating superior efficacy in addressing dataset imbalance and enhancing robustness in dynamic real-time IoT scenarios.