A DeepSHAP-Based Adversarial Attack on Machine Learning-Based Network Intrusion Detection

Many machine learning-based approaches have been developed for network intrusion detection to identify malicious activities and protect network infrastructures. However, maintaining high detection accuracy and robustness remains a significant challenge due to evolving cyber threats, adversarial attacks, and the increasing complexity of network traffic. In this work, we present a DeepSHAP-based adversarial attack framework for evaluating the robustness of machine learning-based Network Intrusion Detection Systems (NIDS). The proposed methodology employs Cascade Forest, Probabilistic Neural Network, and Echo State Network to learn and classify network traffic patterns into normal and attack categories. DeepSHAP is utilized to identify influential features contributing to model predictions and generate adversarial examples that manipulate these critical features. The generated adversarial samples are used to evaluate the vulnerability and robustness of the intrusion detection models. Experimental results demonstrate the impact of adversarial attacks on detection performance and provide insights into the resilience of different machine learning architectures. The proposed framework contributes to improving the security and reliability of intrusion detection systems by enabling comprehensive robustness assessment and supporting the development of more secure cyber-defense mechanisms.

Keywords: Network Intrusion Detection System (NIDS), DeepSHAP, Adversarial Attack, Cascade Forest, Probabilistic Neural Network, Echo State Network, Explainable Artificial Intelligence (XAI), Machine Learning, Cybersecurity, Network Security.