Hybrid Optimized Dual-Function Logistic Regression-based Ensemble Architecture for Robust Password Strength Prediction

Abstract

The increasing frequency of cyberattacks and data breaches has made password strength a critical problem to predict. While various Machine Learning (ML) methods have been applied to password strength classification, their performance is often compromised by class imbalance and the inability to tap into complementary model behaviors. In this paper, based on the Password Security Sber Dataset, this study introduces a novel heterogeneous ensemble method with a weighted soft-voting strategy leveraging K-Nearest Neighbors, XGB, CatBoost, and Logistic Regression. Among the most valuable new contributions is the dual utilization of Logistic Regression as a base classifier and a surrogate optimizer in optimizing ensemble weights dynamically to ensure the system's reliability and stability. The methodological pipeline consists of SMOTE-based oversampling for class imbalance handling, feature selection to preserve discriminative password features, and structured hyperparameter tuning for each base learner. LR optimization is incorporated in the ensemble system for regulating weight assignment during soft voting application for optimal predictive performance. Experimental results show that LR alone achieved 98.45% accuracy and 97.53% F1-score, while the optimized ensemble achieved 98.24% accuracy, 97.21% F1-score, and 98.82% precision. Compared to baseline ensembles and traditional models, the new approach demonstrates improved accuracy, generalizability, and computational complexity and marks its pragmatic significance in providing more robust password strength estimation for modern digital systems.