A Comprehensive Survey of Deep Learning-Based Image Steganalysis: From Statistical Methods to Adversarial Robustness

This survey presents an overview of the development of image steganalysis from basic statistical models to the latest deep learning architectures. A total of 24 significant studies were analyzed, which are grouped into six important dimensions: embedding domain coverage, dataset diversity, low-payload sensitivity, adversarial robustness, pixel-level defense evaluation, computational efficiency. It also shows that the literature is fragmented, with no existing model that effectively covers the cases of heterogeneous dual-domain detection, realistic adversarial robustness on pixel level, and lightweight deployment. Three main gaps were identified: cover-source mismatch in the case of homogeneous training data, single domain architectural limitations, feature space adversarial evaluation that does not represent actual threat models; a unified taxonomy for future research is proposed. This survey provides a well-founded base for quantifying dataset heterogeneity from an information theoretic perspective, providing a principled ground for assessing generalisability in steganalysis systems.