ADV-Sword: A Framework of Explainable AI-Guided Adversarial Samples Generation for Benchmarking ML-Based Intrusion Detection Systems

An expanding number of systems and gadgets connected to the Internet have made themselves easy targets for hackers due to the rapid advancement of technology. Hence, Intrusion Detection Systems (IDS) are essential for protecting network security by identifying cyberattacks in response to these growing threats. Even though machine learning/deep learning (ML/DL) is widely used in intrusion detection systems (IDS), it is still vulnerable to adversarial attacks. These assaults are derived from the initial attack samples and aim to misclassify the judgments made by the models. In this paper, we present ADV-Sword, a novel method that creates resilient adversarial samples by utilizing Accured Malicious Magnitude (AMM) and SHapley Additive exPlanations (SHAP). ADV-Sword is made to go beyond IDS defenses by taking advantage of flaws in ML-based detection systems. Our system comprises two stages, including manipulating malicious features to be benign-oriented and attacking the target IDS models. Using the InSDN dataset, we evaluate ADV-Sword and demonstrate how it can greatly affect IDS performance. According to our test results, ADV-Sword can achieve significant evasion rates, which highlights the detection of IDS at nearly zero.