Abstract:
The integrity and security of computer networks are jeopardized by network intrusions, which
can result in fraudulent activities and illicit access. It interferes the network operations, which
may have disastrous repercussions, including the disclosure of highly confidential data, harm
to one’s reputation, and financial losses. To defend computer networks effective network detec tion and mitigation solutions are essential to lowering the effects of these hazards and ensuring
the overall security of network infrastructures. In this study, we concentrated on improving
network intrusion detection system performance using the Mqtt-IoT-IDS2020 dataset. To re solve the issues presented by unbalanced data, we used two alternative advanced data balancing
methodologies, a chunk-based technique to create subsets of data and a Generative Adversarial
Network (GAN) to create synthetic data. In supervised learning to evaluate and compare these
approaches, we used a different machine and deep learning algorithms in which Random forest
(RF), Decision Tree (DT), and LSTM performed outstanding achieving almost 98% accuracy
in many cases. Furthermore, the problem of unlabeled data is resolved in this study with the
help of Flow-based and TCP-based features in the dataset by evaluating through clustering
models K-mean, Deep Gaussian Mixture Model (DGMM), and BIRCH. Using these models
the evaluation matrices Silhouette Score and Davies- Bouldin for being well discrete and sig nificant grouped features BIRCH performed outstanding, for Calinski Harabasz which gives
ratio within-to-between cluster dispersal K-mean performed highest in all dataset types.
