An Explainable and Carbon‐Aware Stacked Ensemble Model for Phishing and Impersonation Detection in Online Social Networks

ABSTRACT This study presents a structured ensemble framework for detecting phishing and impersonation activities in online social networks (OSNs) with an emphasis on accuracy, interpretability, and sustainability. The framework integrates three gradient boosting models—XGBoost, LightGBM, and CatBoost—within a weighted stacking architecture, where logistic regression serves as the meta‐learner to combine probabilistic outputs. To enhance feature relevance and reduce redundancy, Boruta, LASSO, and Recursive Feature Elimination (RFE) are applied independently to each base learner. Model transparency is ensured through SHAP and LIME, enabling both global and instance‐level explanations of predictions. The system incorporates an energy‐aware evaluation module that measures power consumption (Wh) and carbon emissions (CO 2 e) and introduces the Security‐to‐Sustainability Ratio (SSR) to quantify performance relative to environmental cost. Experiments are conducted on a combined dataset of 98,450 instances, including a Twitter phishing dataset and a curated OSN fake business promotion dataset. A leakage‐safe 5‐fold cross‐validation protocol with Optuna‐based hyperparameter optimization ensures reproducibility and statistical reliability. Results show that XGBoost with Boruta achieves the highest accuracy (96.3%), F1‐score (95.4%), and AUC (0.978), while LightGBM with LASSO provides the lowest latency (15.2 ms) and reduced model size (30.1 MB). The ensemble demonstrates consistent performance across evaluation metrics, with SSR values indicating balanced trade‐offs between detection capability and environmental impact. The proposed framework provides a reproducible and interpretable solution for OSN phishing detection, addressing limitations related to feature optimization, explainability, and sustainability within a unified pipeline.