Publikation
An AI-Powered Approach to File Carving in Digital Forensics – Undo Data Loss with Style
Andrey Guzhov; Tobias Wirth; Matthias Müller; Tobias Fischer; Lucas Howes; Dominik Ospelt
In: Proceedings of the 24th European Conference on Cyber Warfare and Security. European Conference on Cyber Warfare and Security (ECCWS-2025), 24th European Conference on Cyber Warfare and Security, located at ECCWS-2025, June 26-27, Kaiserslautern, Germany, Academic Conferences & Publishing International, 2025.
Zusammenfassung
File carving reconstructs deleted or fragmented files without file system metadata, yet scales poorly on large datasets. We propose an AI-driven classification pipeline evaluating two architectures — a lightweight CNN (M1, 290K parameters) and a Swin Transformer (M2, 28.3M parameters) — across 75 file types. Our results demonstrate that transformer-based models are a promising foundation for next-generation forensic recovery tools, with M2 achieving superior Top@1 accuracy, paving the way toward robust, automated file reconstruction even under real-world conditions involving compression artifacts and embedded formats.
