Machine learning-based detection method for malicious PDF files: A temporal classification approach

Doo-Seop Choi

doi:10.1016/j.asoc.2025.114461

Machine learning-based detection method for malicious PDF files: A temporal classification approach

2026-01-21

0

PRE

AI

DOI：10.1016/j.asoc.2025.114461

原文链接

原文求助

摘要

En 中文

Cybercriminals increasingly exploit non-executable files that can bypass antivirus software detection and are often opened by users without suspicion. In particular, PDF files have become a primary attack vector for adversaries due to their platform-independent nature and ability to preserve document components across different systems. Malicious PDF files continuously evolve to avoid detection, and traditional detection methods, which rely primar ily on static features from older PDF datasets, show limitations in identifying evolving malicious PDF files. This paper identifies temporal evolution in feature distributions and proposes a novel framework to detect malicious PDF files by introducing temporal classification and addressing the evolved characteristics of recent threats. Through in-depth statistical analysis, we revealed that recent malicious PDF files closely mimic the structural characteristics of legitimate files, exhibiting an 11-fold increase in graphic components and a 21-fold increase in hyperlinks compared to older samples. This finding indicates a significant shift in attack methodologies from traditional script injection to social engineering techniques. To address this challenge, we enhanced the basic fea ture set, comprising 31 structural and metadata-based features initially defined in the CIC-Evasive-PDFMal2022 dataset, by integrating 12 newly identified features, resulting in an enhanced set of 43 features. Experimental results demonstrate that our framework with the enhanced feature set achieves 97.80 % detection accuracy us ing the random forest algorithm, representing a 4.12 % improvement over the basic feature set. The framework maintains balanced performance across all metrics with a recall of 0.96, a precision of 0.98, an F1-score of 0.97, and an AUC of 0.99. Additionally, the framework reduced the false positive rate (FPR) from 2.84 % to 1.12 %, a 1.72 percentage points reduction, which is critical for practical deployment in real-world security environments. The proposed enhanced feature set provides an effective approach for strengthening real-world detection systems, including email attachment scanners and antivirus engines, against evolving PDF-based attacks.

Keyword:

PDF malware

Non-executable malware

Malware detection

Machine learning

Analysis of temporal feature evolution

Machine learning-based detection method for malicious PDF files: A temporal classification approach

摘要

期刊

机构

引用论文