Privacy-Preserving Feature Engineering Framework for Real-World Vulnerability Risk Prediction

Abstract

This study introduces a privacy-preserving data preparation and feature engineering framework designed for machine learning-based cybersecurity vulnerability risk prediction, utilizing real-world enterprise scan data from an operational cloud environment. To transform raw vulnerability records into a structured, machine-readable format while complying with personal data protection regulations, the framework integrates systematic data cleaning, missing value imputation, categorical encoding, text normalization, and host pseudonymization. Exploratory data analysis (EDA) was conducted on 36,940 operational records to examine dataset characteristics, vulnerability distributions, and severe class imbalances. Multiple classifiers—including Logistic Regression, Random Forest, XGBoost, and LightGBM—were evaluated under extreme imbalance conditions mitigated via SMOTE. Experimental results demonstrate that non-linear tree-based ensembles achieved the highest predictive performance, with LightGBM attaining a macro F1-score of 0.9974. Comparative analysis further indicates that the effectiveness of textual representations depends on the underlying classifier architecture, with semantic embeddings yielding the best performance when combined with tree-based ensemble models, while TF-IDF remained competitive for linear classification. These findings underscore the critical role of structured text preparation and semantic representation in enhancing model reliability, interpretability, and scalability for automated vulnerability prioritization.