MalPDF-EFS: Enhanced Feature Set Dataset for Malicious PDF Detection (Benign, Old, and Recent Malicious Samples)

MalPDF-EFS is a sample of the feature data used in the following article:

D.-S. Choi, T. Kim, B. Kang, and E. G. Im, "Machine learning-based detection method for malicious PDF files: A temporal classification approach," Applied Soft Computing, vol. 189, Art. no. 114461, Mar. 2026 (available online 23 Dec. 2025). https://doi.org/10.1016/j.asoc.2025.114461

If you find this dataset useful in your research, please consider citing the article above. A BibTeX entry is provided in the Citation section. Thank you.

About the paper

Malicious PDF files change over time, and a detector built on older samples may not work as well on newer ones. The paper studies this by grouping malicious PDF files by their creation date, comparing the groups, and proposing an enhanced feature set for recent malicious PDF files.

1. Data collection. Malicious PDF files were collected with a custom web crawler from public malware repositories (Triage, Hybrid Analysis, and Malware Bazaar) in addition to the existing Contagio dataset. Benign PDF files came from Digital Corpora and Google.

2. Creation-date extraction. The creation date of each file was read from the CreateDate attribute in the XML structure of the PDF. If that failed, the CreationDate attribute in the PDF metadata was used instead. ElementTree was used for XML parsing and PyMuPDF for metadata analysis.

3. Grouping. Malicious PDF files created between 2020 and 2024 were marked as recent, and those created before 2020 as old. Together with the benign files, this gives three groups: Benign, Old malicious, and Recent malicious.

4. Feature sets. The basic feature set has 31 features (general and structural features initially defined in CIC-Evasive-PDFMal2022). After analyzing how the features of old and recent malicious PDF files differ, the paper adds 12 new features, giving the enhanced feature set of 43 features.

5. Results in brief (Random Forest, as reported in the paper on the full dataset):

Experiment	Feature set	Accuracy
Benign vs. old malicious (CASE I)	Basic (31)	99.02%
Benign vs. recent malicious (CASE II)	Basic (31)	93.69%
Trained on old, tested on recent malicious (CASE III)	Basic (31)	87.97%
Benign vs. recent malicious	Enhanced (43)	97.80%

With the basic feature set, accuracy dropped for recent malicious PDF files, and a model trained on old malicious files missed many recent ones (recall 0.7420 in CASE III). The enhanced feature set raised accuracy on recent malicious PDF files to 97.80%. Please see the article for the full results of all models.

About this dataset

This repository releases a sample of the experimental data of the paper: 1,000 randomly sampled PDF files per group, 3,000 in total. It is not the full dataset, so the numbers reported in the paper cannot be reproduced exactly from these files. The remaining data are available from the authors on request.

Only features extracted from the PDF files are included. The PDF files themselves, file names, and file hashes are not included. The features are static features of each PDF file; time is reflected only in the old / recent grouping of the malicious files.

Files

Folder	File	Rows	Columns
FS31/	feature31_benign.csv	1,000	31 features + class
FS31/	feature31_old_malicious.csv	1,000	31 features + class
FS31/	feature31_recent_malicious.csv	1,000	31 features + class
FS43/	feature43_benign.csv	1,000	43 features + class
FS43/	feature43_old_malicious.csv	1,000	43 features + class
FS43/	feature43_recent_malicious.csv	1,000	43 features + class

• FS31 is the basic feature set (Table 2 of the paper). FS43 adds the 12 features proposed in the paper (Table 8). Row i of an FS31 file and row i of the matching FS43 file describe the same PDF.
• class is benign, old_malicious (created before 2020), or recent_malicious (created in 2020-2024).

Features

Basic 31: pdfsize, metadata size, pages, xref Length, title characters, isEncrypted, embedded files, images, text, header, obj, endobj, stream, endstream, xref, trailer, startxref, Page, encrypt, ObjStm, JS, JavaScript, AA, OpenAction, AcroForm, JBIG2Decode, RichMedia, Launch, EmbeddedFile, XFA, Colors. This is the basic feature set of the paper. It follows the feature definitions of CIC-Evasive-PDFMal2022; please also cite that dataset when you use these features.

Additional 12 (enhanced feature set):

Column	Name in the paper	Meaning
hyperlink	hyperlink	number of hyperlinks
entropy	file entropy	entropy of the PDF byte sequence
creator	creator	number of characters in the metadata creator field
producer	producer	number of characters in the metadata producer field
author	author	number of characters in the metadata author field
layout size	page layout size	page area (width x height)
creationDate	creationDate	1 if a valid creation date is recorded, else 0
modDate	modDate	1 if a valid modification date is recorded, else 0
font	font	number of fonts
email	email	number of email addresses
image size	embedded image size	total size of embedded images (bytes)
graphics	graphics	number of vector graphic objects

Notes: values are the raw extractor output. pdfsize and image size are in bytes (the paper reports KB), and layout size is not log-scaled (the paper reports the log10 value). header is the PDF version string and is the only non-numeric feature.

Sampling

Rows whose 31 basic features were identical to another row were removed first, then 1,000 rows per group were drawn uniformly at random with a fixed seed (42).

Citation

If you use this dataset, please cite:

@article{choi2026temporal,
  author  = {Choi, Doo-Seop and Kim, Taeguen and Kang, BooJoong and Im, Eul Gyu},
  title   = {Machine learning-based detection method for malicious {PDF} files: A temporal classification approach},
  journal = {Applied Soft Computing},
  volume  = {189},
  pages   = {114461},
  year    = {2026},
  month   = mar,
  note    = {Available online 23 December 2025},
  doi     = {10.1016/j.asoc.2025.114461}
}

The dataset itself is archived on Zenodo (https://doi.org/10.5281/zenodo.23356149). If you would also like to refer to the dataset directly:

@dataset{choi2026malpdfefs,
  author    = {Choi, Doo-Seop and Kim, Taeguen and Kang, BooJoong and Im, Eul Gyu},
  title     = {{MalPDF-EFS}: Enhanced Feature Set Dataset for Malicious {PDF} Detection (Benign, Old, and Recent Malicious Samples)},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.23356149}
}

License

Creative Commons Attribution 4.0 International (CC BY 4.0).