RISK-AWARE CYBERATTACK DETECTION IN WIRELESS NETWORKS CONSIDERING THE ROBUSTNESS OF MACHINE LEARNING MODELS

Abstract

Intrusion detectors for wireless networks are still commonly assessed on unchanged test sets. Such evaluation does not cover a deliberate modification of controllable traffic features or the use of model explanations to identify a vulnerable direction. A high test score therefore says little about detector behaviour under an active attack. This paper examines a risk-aware pipeline that combines a gradient-boosted base classifier, SHAP explanations, a logistic risk meta-classifier, and selective referral to a human analyst. Risk is described by three signals: the base attack probability, cosine stability of a local SHAP explanation relative to a class-specific training centroid, and Mahalanobis distance from the training distribution. This combination is intended to capture not only the predicted class, but also whether the reasoning pattern and the record itself are familiar to the model. A fixed five percent of records is referred for review according to explanation instability, predictive uncertainty, or predictive entropy. The experiments use a controlled synthetic dataset and NSL-KDD. Two attacks are evaluated: a one-step, gradient-free SHAP-sign perturbation and a transferred projected-gradient attack. The results are not uniform across datasets. On the synthetic benchmark, uncertainty and entropy provide the best selective results; under transferred PGD, macro-F1 on the retained 95 percent rises from 0.882 ± 0.008 for the base detector to 0.908 ± 0.008. NSL-KDD gives a different ordering. Explanation stability is more useful on clean and SHAP-sign-perturbed records, whereas under transferred PGD neither selective configuration improves on C3: macro-F1 is 0.699 for C3, 0.696 for C4, and 0.694 for C4-conf/ent. The base detector also deteriorates markedly, from 0.791 ± 0.003 on clean NSL-KDD to 0.579 ± 0.004 under SHAP-sign and 0.637 ± 0.002 under transferred PGD. These results do not support choosing a review rule from clean validation data alone. Selective referral can manage the accuracy–coverage trade-off, but it is not an adversarial defence. Before deployment, the pipeline should be tested on current Wi-Fi, IoT, 5G, and 6G traffic, with semantically valid perturbations, adaptive attacks, and review rates matched to analyst capacity.