Semantics-Aware Adversarial Training Using Convolutional Neural Networks for Cybersecurity Intrusion Detection
Kavitha Thiyagarajan, Ali Alkwzahy, B S Shruthi, S. Gayathri, S Kiruthuka
Abstract
Nowadays, the Convolutional Neural Networks (CNNs) have been widely employed for network intrusion detection due to their ability to capture spatial and sequential traffic patterns. However, existing intrusion detection systems often fail when subjected to adversarial perturbations, where attackers slightly manipulate the traffic features to evade detection. This vulnerability highlights the pressing challenge of building adversarially robust cybersecurity models. Hence, this research proposes Semantics-Aware Adversarial Training using CNN (SAAT-CNN) for intrusion detection. Initially, the input data were collected from the CICIDS2017 dataset, which contains realistic benign and malicious traffic. Then, the collected data are preprocessed to handle the missing values, numeric features are normalized using Standardscaler, label encoding is employed to transform the categorical features, and domain constraints are defined to preserve feature semantics. Furthermore, a <tex xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">$1 \mathrm{D}-\text{CNN}$</tex> is trained for initial intrusion classification, and adversarial examples are then generated using the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD) within domain-preserving bounds. Finally, adversarial training was employed on the mixed clean and perturbed data to minimize false negatives. The proposed SAAT-CNN provided better results in terms of detection accuracy (99.99 %) when compared to the existing 1D-CNN model.