A predictive and reinforcement learning-driven framework for adaptive cybersecurity training and behavioral risk mitigation
Phishing attacks continue to become more sophisticated and employee behavior is still a major part of the organization’s cyber security strategy. This study introduces a behavior-aware framework to predict the phishing susceptibility through supervised learning, clustering (unsupervised learning techniques), and reinforcement learning and to provide the customized cybersecurity training. The framework was created based on an augmented dataset, which takes into account phishing simulation data along with Human Resource (HR) data, such as demographic, behavioral, and organizational attributes. The accuracy, F1 Score, ROC-AUC, and confusion matrix were used to assess the accuracy of the five ensemble learning algorithms (Random Forest (RF), XGBoost (XGB), Extra Trees (ET), LightGBM (LGBM) and soft Voting Ensemble). Among the baseline classifiers, ET outperforms as compared to other by achieving accuracy (80%), and F1-score (79%) with ROC-AUC (0.88) which was good for discriminative power over decision thresholds. To identify various behavioral personas of employee ‘s behaviour, K-Means clustering algorithm was applied. The optimal clustering solution was determined using silhouette analysis, which was followed by one-way ANOVA that revealed a significant difference in the groups’ phishing risk and their engagement and participation in training. These behavioral profiles create interpretable basis to create specific cybersecurity interventions. The framework is designed to build on these insights and uses Q-learning to adapt training approaches based on the user’s behavior response instead of using fixed training programs based on awareness. To the best of our knowledge, this is among the first framework to integrate phishing susceptibility prediction, behavioral segmentation, and adaptive based reinforcement training within unified human-centered cybersecurity architecture.
Authors
- Samar M. Alqhtani (ORCID: https://orcid.org/0000-0002-8664-8953)
- Mohammad Alsulami (ORCID: https://orcid.org/0000-0001-5765-1291)
- Mana Saleh Al Reshan (ORCID: https://orcid.org/0000-0002-2266-9608)
- Darakhshan Syed (ORCID: https://orcid.org/0000-0002-5319-5847)
- Hamad Almansour (ORCID: https://orcid.org/0009-0007-4263-2881)
- Asadullah Shaikh
- Mehak Samoo
Institutions
- Isra University (PK)
- Najran University (SA)
- Bahria University (PK)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-21
- DOI
- https://doi.org/10.1038/s41598-026-81813-2
- Primary Topic
- Spam and Phishing Detection
- Type
- article
- Field-Weighted Citation Impact
- 0.00