An interpretable approach for trustworthy intrusion detection systems against evasion samples

Abstract

In recent years, Deep Neural Networks (DNN) have demonstrated remarkable success in various domains, including Intrusion Detection Systems (IDS). The ability of DNN to learn complex patterns from large datasets has significantly improved IDS performance, leading to more accurate and efficient threat detection. Despite their effectiveness, DNN models exhibit vulnerabilities to adversarial attacks, where malicious inputs are specifically crafted to deceive the models and evade detection. This paper provides insights into the effectiveness of deep learning-based IDS (DL-IDS) against adversarial example (AE) attacks. We tackle the weaknesses of DNN in detecting adversarial attacks by proposing the Convolutional Neural Network (CNN), which serves as an AE detector. We also utilize one of the XAI techniques, specifically SHAP, to enhance the transparency of the AE detector. Our results show that the AE detector has obvious effects for detecting adversarial examples and achieves an impressive 99.46% accuracy in our experimental environment.