Skip to content
Open access

Source Identification of OOXML Documents Using Internal Compression Structure Patterns for Digital Forensic Investigation

2026 · IEEE Access · Vol 14, pp. 161359-161376 · 0 citations · 34 references

Abstract

Identifying the program that last saved an electronic document is critical for linking digital evidence to a suspect’s working environment in forensic investigations. The conventional approach inspects the Application tag in the app.xml metadata of Office Open XML (OOXML) documents, but several editors store identical tag values, omit the tag, or do not generate the metadata part at all, and the metadata itself can be trivially modified or removed as an anti-forensic measure. This paper identifies and systematically validates that each document editor produces a distinctive ordering of compressed parts when saving an OOXML file, and leverages this ordering as a sequential feature for source identification. Using a dataset of 19,500 documents generated by 13 document editing programs across 15 distinct editors, we compare 15 classification models under identical preprocessing and data-split conditions. The best models achieve 99.42–99.93% accuracy across Word, PowerPoint, and Excel documents, whereas even the most favorable construction of the metadata-based approach reaches an expected accuracy of only 66.7–75.0% and becomes inapplicable to reliable source attribution once the original metadata is deliberately removed. Ablation, recompression, and cross-version analyses further show that removing the original ordering reduces the best observed accuracy by 22–37 percentage points, that archive-level reordering can substantially reduce classification confidence and motivates a confidence-based triage deployment, and that the trained models correctly attribute all 190 cross-version samples without retraining. A forensic case study illustrates how the classification results can be integrated with operating system artifacts to trace a document’s lifecycle on a suspect’s system. The preprocessed dataset and a standalone classification tool are publicly available.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.