Structured Exponentiation Dot Product Attention-based Vision Transformer with Coordinated FuseNet Transfer Learning for Network Intrusion Detection

Abstract

The number and intensity of cyberattacks have grown exponentially with the explosive growth of cloud computing,
Internet of Things (IoT), and high-speed networks. According to recent reports, more than 70% of the contemporary networks intrusions are either stealthy or low-rate intrusions, which are proven to be hard to detect with the traditional Intrusion Detection Systems (IDSs). Despite their extensive implementations in signature-based and rule-based IDS, these methods have weaknesses in terms of being slow to adapt to zero-day attacks, a high number of false alarms, and a lack of scalability when dealing with vast data volumes and high-dimensional network traffic data. To overcome these issues, this research proposes a Structured Exponentiation Dot product-based Network Intrusion Detection System (SED-NIDS) to facilitate the determination of accurate and robust cyberattacks. The proposed system is end-to-end, unlike single-model IDS methods, which perform feature learning and decision fusion independently, leading to better generalization of the system to various attack scenarios. The proposed SEDNIDS achieves an accuracy of 98.6%, 93% precision, 94% recall, and an F1-score of 93.5%. The proposed system also displays a reduced false alarm rate of 1.3% indicating how the system is applicable in real-time deployment. SED-NIDS is a precise and robust IDS that will answer the changing demands of the contemporary cyber threats. The proposed model has great improvements in detection performance with low false positives which makes it a viable solution to the next-generation network security systems.