Ensemble learning is a strategy commonly used to fuse different base models by creating a model ensemble that is expected more accurate on unseen data than the base models. This study describes a new cyber-threat detection method, called , that uses ensemble learning coupled with adversarial training in deep learning, in order to gain accuracy with neural models trained in cybersecurity problems. The selection of the base models is one of the main challenges to handle, in order to train accurate ensembles. This study describes a model ensemble pruning approach based on eXplainable AI (XAI) to increase the ensemble diversity and gain accuracy in ensemble classification. We base on the idea that being able to identify base models that give relevance to different input feature sub-spaces may help in improving the accuracy of an ensemble trained to recognise different signatures of different cyber-attack patterns. To this purpose, we use a global XAI technique to measure the ensemble model diversity with respect to the effect of the input features on the accuracy of the base neural models combined in the ensemble. Experiments carried out on four benchmark cybersecurity datasets (three network intrusion detection datasets and one malware detection dataset) show the beneficial effects of the proposed combination of adversarial training, ensemble learning and XAI on the accuracy of multi-class classifications of cyber-data achieved by the neural model ensemble.

Cyber-attacks are increasingly becoming a major concern for individuals and organizations alike. Meanwhile, attackers are employing advanced techniques taking advantage of the growing power of Artificial Intelligence (AI) to develop highly sophisticated attacks at an accelerated pace. Consequently, developing effective tools to detect cyber-attacks and protect digital assets has become of utmost importance to practice and research. Numerous AI-based techniques, mainly Machine Learning (ML) and Deep Learning (DL), have been investigated in the literature to protect digital assets from cyber threats. Particularly, DL has received significant attention in recent cybersecurity research due to its powerful capabilities in managing huge amounts of data and detecting malicious cyber threats. The majority of the proposed threat-detecting techniques in the literature focus on using Deep Neural Networks (DNNs). However, some research in the literature utilized the capabilities of Convolutional Neural Networks (CNNs) in detecting cyber threats by converting the cybersecurity data from tabular data into images. This paper aims to investigate the efficiency of both approaches in detecting cyber threats with the help of eXplainable AI (XAI). Our findings indicate that on NSL-KDD, the DL model trained on tabular features achieved WeightedF1 = 0.74, MacroF1 = 0.56, and Overall Accuracy (OA) = 0.78, compared with 0.73, 0.54, and 0.76 for the image-based model. On CICMaldoid20, the tabular model achieved 0.79 (WeightedF1), 0.80 (MacroF1), and 0.80 (OA), compared with 0.76, 0.75, and 0.77 for the image-based counterpart. These results suggest that while image transformations can be beneficial in specific classes, tabular models consistently deliver stronger overall performance. Received: 25 June 2025 | Revised: 12 November 2025 | Accepted: 21 November 2025 Conflicts of Interest The authors declare that they have no conflicts of interest to this work. Data Availability Statement Data are available from the corresponding author upon reasonable request. Author Contribution Statement Malik AL-Essa: Conceptualization, Methodology, Software, Validation, Formal analysis, Resources, Data curation, Writing – original draft, Writing – review & editing. Mohammad Alsharo: Conceptualization, Methodology, Software, Validation, Formal analysis, Investigation, Resources, Data curation, Writing – review & editing. Yazan Alnsour: Conceptualization, Methodology, Software, Validation, Resources, Data curation, Writing – review & editing. Wasim A. Ali: Conceptualization, Methodology, Software, Validation, Formal analysis, Resources, Data curation, Writing – review & editing. Omar Almomani: Writing – review & editing, Project administration.

Machine Learning (ML) intrusion detection systems (IDS) are vulnerable to manipulations: small, protocol-valid manipulations can push samples across brittle decision boundaries. We study two complementary remedies that reshape the learner in distinct ways. Adversarial Training (AT) exposes the model to worst-case, in-threat perturbations during learning to thicken local margins; Counterfactual Augmentation (CF-Aug) adds near-boundary exemplars that are explicitly constrained to be feasible, causally consistent, and operationally meaningful for defenders. The main goal of this work is to investigate and compare how AT and CF-Aug can reshape the decision surface of the IDS. eXplainable Artificial Intelligence (XAI) is used to analyze the shifts in global feature importance stability under both AT and CF perturbation to link these shifts to the accuracy of the IDS in detecting cyber-threats. This yields a clear picture when boundary hardening (AT) or boundary sculpting (CF-Aug) better serves IDS. Two well-known techniques are used to generate adversarial samples, namely the Fast Gradient Sign Method (FGSM) and the Projected Gradient Descent (PGD) techniques. We have achieved better accuracy with AT and CF-Aug compared to the baseline IDS.