Attention-based CNN for Adversarial File Fragment Detection Against Padding and Bit-Flip Attacks

Title

Attention-based CNN for Adversarial File Fragment Detection Against Padding and Bit-Flip Attacks

Creator

Mary, Teena; Sreeja, C.S.

Description

File fragment classification represents a critical task within digital forensics and cybersecurity that aims to recover fragmented files when their metadata is not available. Even though cutting-edge deep learning models achieve 77-79% accuracy on clean fragments, none of the existing file fragment classification systems currently include detection mechanisms against adversarial attacks, thus remaining defenseless against attackers using byte-level perturbations. This paper addresses this gap by proposing the first adversarial detection framework for file fragment classification. This paper presents an attention-based CNN that combines byte embeddings with both spatial and channel attention mechanisms to detect byte-level perturbations before actual classification. Evaluated over 30.72 million fragments across 75 file types, the detector reaches an accuracy of 91.44% against five attack strategies: null-byte padding, random-byte padding, cross-file padding, random bit-flipping, and header-targeted bit-flipping, at 91.34% recall, 95.46% specificity, and 0.9819 AUC-ROC. With 1.31 M parameters and 1 ms inference time per fragment, the detector enables practical deployment as a preprocessing filter within two-stage forensic pipelines screening suspicious fragments before reaching standard classifiers. This foundational work sets up the first comprehensive benchmark for adversarial robustness evaluation specifically in file fragment classification. 2025 IEEE.

Source

Proceedings of 5th International Conference on Evolutionary Computing and Mobile Sustainable Networks, ICECMSN 2025;pp.935-942

Date

01-01-2025

Publisher

Institute of Electrical and Electronics Engineers Inc.

Subject

adversarial detection; attention mechanism; cybersecurity; deep learning; digital forensics; file fragment classification

Coverage

Mary T., Christ University, Department of Computer Science, Bengaluru, India; Sreeja C.S., Christ University, Department of Computer Science, Bengaluru, India

Rights

Restricted Access; Hardcopy may be available in the library

Relation

ISBN: 979-833158242-5;

Format

online

Language

English

Type

Conference paper

Identifier

https://doi.org/10.1109/ICECMSN68058.2025.11382778

https://www.scopus.com/pages/publications/105035543391?origin=resultslist