Adversarial-Aware Deep Learning Frameworks for Enhanced Network Threat Detection

Abstract

Network Intrusion Detection Systems (NIDS) are used as the first line of defense against cyber threats since they protect vital data and network resources. Nonetheless, when Deep Learning (DL) is applied to NIDS, they are very susceptible to adversarial attacks, which interfere with their reliability. The paper describes a more improved network threat detection model with the CICIDS 2017 dataset, which includes both benign and malicious communication from different protocols. To make the dataset quality and robust, data cleansing, normalization, feature selection, and balancing of classes were performed using SMOTE. An LSTM model was suggested as it has the ability to exploit sequential and temporal relationships in network traffic. The performance of the LSTM was compared to that of the Support Vector Machine (SVM), the Random Forest (RF), and the Deep Neural Network (DNN). 96.5% F1-score, 97% accuracy, 97% precision, and 96% recall were achieved by the LSTM, outperforming the other models, according to the experiment's results. It has been proven through these results that LSTM is an excellent method for the detection of assault patterns that are in flux, which have better detection rates and lower false alarms, rendering it a potentially useful tool in improving network security.