Adversarial Shadows in Digital Forensics: New Insights Into File Fragment Classification Vulnerabilities and Defenses

The paper is a comprehensive survey of adversarial attacks on file fragment classification (FFC) models -a relatively unexplored area in digital forensics, given the increasing application of machine learning techniques. Unlike image or text classification adversarial attacks, adversarial attacks on FFC exploit statistical and structural properties at the byte level in systems that lack semantic or perceptual knowledge. Such properties necessitate the use of domain-specific defense strategies, as the defense strategies adopted from other domains are typically not effective for the problems of FFC. The survey comprehensively evaluates attack mechanisms relevant to FFC, including evasion and poisoning attacks, and discusses their impact on forensic reliability. It highlights the absence of domain-specific benchmarks, robust evaluation protocols, and systematic research on the adversarial robustness of FFC. The paper also discusses the different types of byte level perturbations that can happen in fragment data, and it sets specific research priorities for raising the reliability of machine learning-based digital evidence recovery and security. The paper provides building blocks for future work, offering practical insights for development in ensuring file fragment classification systems utilized in forensics are secure.