Leveraging Explainable AI for Adaptive Adversarial DoS Attack Detection in 6G IoT Networks

The advent of 6th Generation(6G) networks and the rapid expansion of the Internet of Things (IoT) have heightened the security risk potential of such systems to cyberattacks, particularly Denial of Service (DoS) attacks. The 6G networks strongly rely on Artificial Intelligence(AI) for intrusion detection, which yields promising results. In parallel, attackers advance their attacking strategies dynamically, declining the AI-based DoS attack detection accuracy, thereby opening the network for potential attackers flagged as false negatives. This phenomenon, called concept drift, requires dynamic adaptation of the AI-based intrusion detection system using the latest attack data. Explainable AI(XAI) is one of the practical techniques that in-terprets the Machine Learning(ML) model outcome with feature contribution. This paper proposes a novel dynamic adaptation framework for 6G network intrusion detection, integrating real-time monitoring and identifying concept drift from adaptive adversaries using the Wilcoxon stat test. Our work leverages XAI-based feature importance interpretation to optimize required model retraining selectively. This work incorporated Shapley value-based local explanations for selected network traffic samples in the experiment with the UNSW-NB15 attack dataset. We comprehensively evaluated our work with supervised learning techniques for intrusion detection. The results reflect the marginal improvement of DoS classification accuracy by 4.49 % up to 93% in combination with Extreme Gradient Boost(XGB) while marginally reducing training and prediction time by 57.6 % and 47.7% with Gradient Boost(GB) for DoS attack detection beyond state of the art. The results showcase the effectiveness of our work for adaptive adversarial detection in 6G IoT networks with time-efficient dynamic retraining against concept drift with feature selection(19 and 21 out of 42) with improved accuracy.