MODIFIED AQUILA OPTIMISER-BASED INTELLIGENT MODELS FOR EFFECTIVE CLASSIFICATION OF MULTIPLE ATTACKS IN INTERNET OF MEDICAL THINGS

The prevalence rate of sophisticated, financially-motivated and systemic cyberattacks in Internet of Medical Things (IoMT) has adversely affected not only the privacy and availability of healthcare records but also healthcare organisation reputation. IoMT has revolutionised the medical field by making healthcare services more accessible, robust and personalised. Due to its high complex architecture, diverse and heterogeneous nature of different medical devices sharing communication with different protocols, the network is susceptible to different forms of cyberattacks such as data breaches, Denial of service (DoS), Distributed Denial of service (DDos) and so on. This study is aimed at developing improved machine learning (ML)-based multiclass attack classification models in IoMT environment. The specific objectives are to: perform exploratory data analysis and visualisation of the labelled IoMT dataset; carry out data pre-processing to solve the identified open problems in the real-life dataset during exploratory analysis; leverage on exploitation and exploration capability of Aquila optimiser (AO) in selecting relevant and promising features in the dataset while solving premature convergence drawback of AO with the integration of levy flight search to modify AO. The modified AO is used to optimise the feature selection process from the dataset. Different number of features were selected in AO and modified AO experimentations. The extensive data cleaning carried out, helped improve the quality of the data which were later used to build robust machine learning-based (ML) multiclass attack classification models in CICIoMT 2024 dataset using Random Forest, Adaboost and Hidden Naïve Bayes algorithms. Synthetic Minority Over Sampling Technique (SMOTE) was also employed to attend to the extreme class imbalance issue that was identified in the dataset during the EDA. The study equally tuned the hyperparameters of the learning algorithms while classifying the attacks. In addition, stratified K-fold cross validation was adopted to ensure reliability of the developed models. The methodological approach employed in this study which is a multi-stage learning approach emphasises achieving ML models with promising classification performance with lesser false positive rate. Equally, the best performance model was selected and deployed. The experimental results showed that Random Forest hybridized with AO achieve accuracy of 99.69%, precision of 94.40%, recall of 95.54%, f1-score value of 94.93% and ROC-AUC value of 99.58% respectively. However, modified AO achieved higher accuracy of 99.74%, precision of 0.9975, recall of 0.9974, f1 score value of 0.9975 and ROC-AUC value of 0.9994 were recorded. Adaboost hybridized with modified AO achieving second highest accuracy of 98.31%, precision of 85.52%, recall of 94.93%, f1-score value of 86.33% and ROC-AUC value of 86.33%. These results in the modified AO did not only outperform state-of-the-art methods but also indicate the efficiency of the proposed method in identifying a multiple of attacks in IoMT landscape. It is concluded that, the proposed modified AO framework represents a significant advancement in securing IoMT environments through intelligent, robust, and computational efficient multiclass attack classification. Further study can explore the integration of hybridized modified AO techniques for identifying attacks in IoMT so as to improve on data privacy and availability.