IMFOHDL-ID: A Hybrid Meta-Heuristic and Deep Learning Framework for Robust Intrusion Detection in IoT Networks

Abstract

The Internet of Things (IoT) has greatly broadened the cyber-attack surface and the inability of traditional Intrusion Detection Systems (IDS) to keep up with the number, complexity and zero-day nature of today's cyber threats. To mitigate the above-mentioned vulnerabilities this study proposes the Improved Mayfly Optimization (IMFO), a novel hybrid intelligence-based system that is specially designed for detecting robust cyberattacks in IoT environments. The proposed methodology has four core algorithmic stages (preprocessing, selecting feature, sequential learning and hyperparameter tuning) implemented across in seven architectural blocks including environment of IOT, Collection, Preprocessing, IMFO, Engine system, intrusion classification and evaluation system. These can be achieved by four stages, which are synergistic to each other: (1) Min-Max normalization for traffic preprocessing; (2) IMFO algorithm for intelligent dimensionality reduction and feature selection; (3) A Long Short-Term Memory-based Deep Stacked Sequence-to-Sequence Auto encoder (LSTM-DSSAE) for learning temporal dependencies and identifying malicious deviations; (4) An autonomous hyperparameter tuning algorithm, called Dipper-Throated Optimization Algorithm (DTOA). The IMFOHDL-ID framework was found to be more effective than baseline methods in terms of accuracy (98.31%), precision (92.09%), recall (91.75%), F-score (91.90%) and Matthews correlation coefficient (MCC =90.96%) on the BoT-IoT dataset. The results show that the proposed architecture is able to defend against comprehensive attacks on real-world IoT networks with high accuracy, scalability and resource efficiency, while achieving deep sequence learning through advanced meta-heuristic optimization.