Ensemble Learning for Proactive Detection of Network-Intrusion-Based Insurance Fraud

Abstract

We propose an ensemble learning pipeline that proactively integrates Stacking Feature Embedding (SFE) with Principal Component Analysis (PCA) and tree-based ensembles to proactively detect insurance fraud originating from network intrusions. The main contributions are: (1) the novel integration of SFE-PCA as a meta-feature construction step for tabular network flow data; (2) a sensitivity analysis that justifies PCA reduction ratios used for each dataset; and (3) a computational and ethical assessment for real-world deployment. Random Forest (RF), Extra Trees (ET), and XGBoost classifiers were trained and evaluated on benchmark intrusion datasets, specifically NSL-KDD, LYCOS-IDS2017, and CIC-IDS2018. Findings from experiments conducted on these datasets show that the proposed pipeline achieves high detection performance (AUC > 0.995) and 99.9% accuracy, while reducing feature dimensionality and resource use compared to deep baselines (CNN/LSTM). These results suggest the approach is an efficient, interpretable option for proactive intrusion-driven insurance fraud detection.