Evaluating and Improving Adversarial Robustness of Deep Learning -Based Network Intrusion Detector
Network Intrusion Detection Systems (NIDS) are essential for protecting digital infrastructures. Deep learning Networks (DLNs) have shown great promise in detecting network threats, while existing IDS solutions continue to suffer from issues such as high false alarms, low detection rates and reduced false positives. Previous research has highlighted the effectiveness of DLs in NIDS and identified their susceptibility to Network attacks, a comprehensive evaluation and improvement framework is lacking. This study aims to fill that gap by developing a robust evaluation framework and implementing defense mechanisms such as Inception module, feature squeezing, multiple kernel sizes, and gradient masking. The UNSW-NB15 dataset will be used to train and test the models. The result obtained is an enhancement in the robustness of DL-based NIDS, measured by improved accuracy, reduced attack success rates, and better learning curves. Strengthening these systems will significantly enhance the security of digital infrastructures, providing a robust defense against sophisticated cyber threats.
Download Paper