Deep Learning for Image Steganalysis: A Structured Comparative Review of Dataset, Domain, and Robustness Coverage

Abstract

Deep learning has substantially advanced image steganalysis, yet inconsistent evaluations across datasets, embedding domains, payload rates, and threat models obscure the field's readiness for real-world forensic deployment. This structured comparative review analyses 26 deep-learning steganalysis studies (2015–2025) across six standardized dimensions, synthesizing findings into three critical structural gaps. First, adversarial robustness is severely under-addressed: only 19% (5 of 26) of studies evaluate robustness, with just a single study testing pixel-level, norm-bounded perturbations rather than feature-space attacks. Second, single-domain architectural specialization predominates, as no reviewed study evaluates a single unified model across both spatial and frequency embedding domains; limiting multi-domain efforts to separate models or content heterogeneity. Third, cover-source mismatch remains widespread due to dataset homogeneity, with most studies relying solely on single-source benchmarks like BOSSBase 1.01. Ultimately, while individual studies partially address single limitations, none solve all three simultaneously; closing these gaps requires a unified architecture combining multi-source training, dual-domain coverage, and pixel-level adversarial testing, representing an essential open challenge for practical deployment.