Semester of Graduation
Spring 2026
Degree
Master of Science in Computer Science (MSCS)
Department
Department of Computer Science and Engineering
Document Type
Thesis
Abstract
Fragmented file reassembly enables the recovery of digital evidence from storage media. Recently, deep learning has been shown to offer efficiency improvements to this process through the automated identification of individual file fragments by file type. However, the decisions deep learning architectures make are abstract and difficult to interpret, often leading the ubiquitous “black box” interpretation. This raises questions about whether their output is trustworthy enough to support criminal evidence collection. This thesis presents explainability studies using a WindowSHAP approach to interpret the output of MoDiCo, a deep learning architecture that enables file type classification at unprecedented scale. This approach shows MoDiCo uses human-interpretable byte patterns to determine a file fragment’s type, thus enabling detection and correction of data representation bias. The findings presented herein demonstrate that this explainabil- ity approach can be used to produce a more valid and transparent model for real-world forensic investigations.
Date
3-25-2026
Recommended Citation
McCain, Joshua D., "Explaining and Interpreting Byte-Level File Type Classification in Deep Learning Models" (2026). LSU Master's Theses. 6317.
https://repository.lsu.edu/gradschool_theses/6317
Committee Chair
Ghawaly, James M.
LSU Acknowledgement
1
LSU Accessibility Acknowledgment
1