Multi-Texture Fusion with ReliefF Subspace-Weighted Support Vector Machines for Lightweight Defect Classification of Electroluminescence Solar Cell Images

Main Article Content

Priya Pal
Dr. Mukesh Yadav

Abstract

Zero-day attacks exploit vulnerabilities for which no signature exists, so an intrusion detection system (IDS) must recognise them from behaviour alone. Many machine learning IDS studies report near-perfect accuracy, but they usually test on attacks of the same types seen in training, which says little about unseen attacks. This paper presents WSE-XIDS, a weighted stacking ensemble with eXplainable artificial intelligence (XAI) that combines a class-balanced multi-layer perceptron (MLP), an extreme gradient boosting (XGBoost) model and a light gradient boosting machine (LightGBM) through a class-balanced logistic regression meta-learner, and explains its decisions with SHapley Additive exPlanations (SHAP). The model is evaluated on the NSL-KDD benchmark under two protocols: a stratified 80/20 split of KDDTrain+ (known attacks) and the official KDDTest+ set, which contains 3,750 records of 17 attack types never seen in training (zero-day test). On known attacks WSE-XIDS reached 99.85% accuracy, 98.31% macro F1-score and the highest macro recall (99.42%) of eight models. On the zero-day test it reached 77.83% accuracy and the best macro F1-score (67.07%), detected 35.81% of the novel-attack records as attacks, and raised false alarms on 5.07% of normal traffic. SHAP analysis shows that source bytes, connection count and destination bytes drive most decisions. The large gap between the two protocols, and the low recall on unseen remote-access attacks, show that zero-day detection remains far from solved, and that known-attack accuracy alone overstates IDS readiness.

Article Details

Section
Articles