Abstract:
With the advancement in internet technology, an increase in network-related threats has
led to several privacy concerns. Automated Intrusion Detection Systems (IDSs) can de tect malicious patterns where firewalls designed using conventional detection techniques fail
to identify the relevant set of threat patterns. There are different types of IDS available,
but those systems are prone to high False Alarm Rate (FAR) because anomalies can be
new legitimate activities. Good quality and surplus network traffic patterns will make IDS
systems more effective. Hence to reduce the FAR and enhance detection accuracy, a novel
method to extract the best features and classify them as ’normal’ and ’intrusive’ is proposed.
Feature extraction is carried out using XGBoost and Autoencoder techniques and classifi cation is performed using TabNet. The proposed method is compared using four classifier
models - XGBoost, Dense Neural Network, Convolutional Neural Network, and Temporal
Convolutional Network. Temporal relations underlying the data are also analyzed. XGBoost
feature extraction is found more efficient for feature extraction when compared to Autoen coder. Also, TabNet exhibited the top performance while comparing with other classifier
models. The experiments are carried out using UNSW-NB15 and NSL-KDD datasets and
the performance is evaluated using other techniques available in the literature
