Lightweight Vulnerability-Aware Malware Triage: Operational Value and Generalization Under Family and Temporal Shift
Year: 2026
Type: article
Abstract: Security operations require effective prioritization of malware samples that warrant deeper investigation. This study evaluates whether five Portable Executable (PE) metadata features—Packed, Entropy, Packer, Compiler, and Linker—can support lightweight vulnerability-aware triage using PEVuln. After hash-based deduplication, 35,800 unique samples remained, including 559 vulnerability-associated samples. Leakage-safe pipelines evaluated a prior baseline, logistic regression, Random Forest, Extra Trees, and LightGBM under random, family-disjoint, and chronological holdout protocols. Hyperparameters were selected by five-fold cross-validation using Average Precision (AP), thresholds on separate validation data using F2, and uncertainty by 1000 stratified-bootstrap repetitions. Under random evaluation, tree ensembles achieved AP of 0.253–0.255, about 16 times prevalence, while a 5% Random Forest review budget recovered 50% of positives. The highest test AP fell to 0.118 under family-disjoint evaluation and 0.015 chronologically. Feature analyses showed strong complementarity: Entropy alone achieved AP 0.0212, yet removing it reduced full-model AP from 0.2553 to 0.1695. A matched cohort comparison demonstrated superior predictive performance for EMBER, while the five-feature representation retained useful ranking capability at a 5% review budget under both random and family-disjoint evaluation. This representation provides useful first-stage ranking signal but limited robustness under family and chronological shift.
DOI: 10.3390/jcp6060201
Institution: Université Sultan Moulay Slimane
Language: English
Subfield: Software
Field: Computer Science
Domain: Physical Sciences
Keywords: temporal shifts, embers
Open Access status: gold
Estimated APC paid: $1,207
OpenAlex ID: W7225893426
