Abstract
Security-audit outputs can support host assessment, but labels produced by an auditing tool should not be interpreted as independently validated risk estimates. This exploratory study examines whether machine-learning models can reproduce Lynis's native warning-versus-suggestion categories for control identifiers not seen during training. Twenty-nine public lynis-report.dat files yielded 1,022 raw findings; after task filtering and exact deduplication, 116 unique observations remained (102 suggestions and 14 warnings). We used a grouped 70/30 hold-out by control ID, training-only class balancing, mutual-information feature selection, and four-fold stratified group cross-validation to compare 12 base classifiers and a 12-model soft-voting ensemble. Gaussian Naive Bayes was selected from cross-validation (macro F1 0.630 ± 0.176) and achieved macro F1 0.536 and balanced accuracy 0.774 on the untouched hold-out. All four test warnings were recovered, but warning precision was 0.222. These results show that the engineered representation contains signal associated with Lynis's existing categories; they do not constitute an independent validation of security-risk prioritization, exploitability, or vulnerability severity. Larger, externally labelled and longitudinal datasets are required before operational risk-prioritization claims can be tested.

This publication has 30 references indexed in Scilit: