Overcoming Dataset Imbalances and Computational Challenges in IoT Intrusion Detection: A SMOTE-Enhanced Transformer-Based Model

Due to security holes in IoT-enabled networks, the massive adoption of IoT has unleashed a new wave of threats into computer networks. Due to their limited resources, these devices are ill-equipped to handle security measures, making them the most vulnerable part of our computer networks and putting our data and systems at risk. Some have suggested Intrusion Detection Systems (IDS) as a means to lessen the impact of intrusions connected to the IoT in an effort to solve this problem. Although IDS are essential for detecting threats, they are not widely used because of their high computing costs and reliance on labelled data. Finding a happy medium between fast detection and excellent accuracy is the objective of this model development effort. In order to accomplish this, to present a feature selection method based on the black-winged kite algorithm (BKA), a groundbreaking swarm intelligence technique that combines the Leader tactic and the Cauchy variation procedure to find the expansive appropriate convergence solution. BKA is based on the migratory and predatory instincts of black-winged kites. The study used few-shot learning techniques by leveraging transfer learning models and specifically using pre-trained architectures for categorisation. It also fine-tuned four pre-trained transformers. To take into account the imbalance in these datasets as a significant component and use the Synthetic Minority Oversampling Technique (SMOTE) to address it. The research team used BoT-IoT, TON-IoT, and CIC-DDoS2019 as benchmark datasets to assess the impact of the chosen features on the suggested model's performance. Metrics for evaluation include F1-score, recall, precision, and accuracy of detection. Evidence from experiments shows that the suggested ensemble model outperforms state-of-the-art AIDS topologies for IoT networks in terms of detection accuracy and efficiency.