Abstract
Devices which are part of the Internet of Things (IoT) have strong connections; they generate and consume data, which necessitates data transfer among various devices. Smart gadgets collect sensitive information, perform critical tasks, make decisions based on indicator information, and connect and interact with one another quickly. Securing this sensitive data is one of the most vital challenges. A Network Intrusion Detection System (IDS) is often used to identify and eliminate malicious packets before they can enter a network. This operation must be done at the fog node because the Internet of Things devices are naturally low-power and do not require significant computational resources. In this same context, we offer a novel intrusion detection model capable of deployment at the fog nodes to detect the undesired traffic towards the IoT devices by leveraging features from the UNSW-NB15 dataset. Before continuing with the training of the models, correlation-based feature extraction is done to weed out the extra information contained within the data. This helps in the development of a model that has a low overall computational load. The Tab transformer model is proposed to perform well on the existing dataset and outperforms the traditional Machine Learning ML models developed as well as the previous efforts made on the same dataset. The Tab transformer model was designed only to be capable of handling continuous data. As a result, the proposed model obtained a performance of 98.35% when it came to classifying normal traffic data from abnormal traffic data. However, the model's performance for predicting attacks involving multiple classes achieved an accuracy of 97.22%. The problem with imbalanced data appears to cause issues with the performance of the underrepresented classes. However, the evaluation results that were given indicated that the proposed model opened new avenues of research on detecting anomalies in fog nodes.