A Taxonomy-Driven Survey of Deep Learning-Based Semantic Mapping Methods for Coverless Steganography

Coverless steganography has emerged as a promising alternative to traditional information hiding techniques by eliminating direct modification of carrier media. Recently, deep learning (DL) has enabled a new paradigm known as deep semantic mapping, where secret information is encoded through high-level semantic representations extracted from images rather than pixel-level embedding. This approach significantly improves resistance against modern steganalysis methods based on statistical and neural network analysis. In this study, we aim to provide a systematic taxonomy and comparative analysis of deep learning-based coverless steganography methods with a focus on semantic mapping techniques. A comprehensive systematic survey of recent literature (2019 onward) is conducted, covering retrieval-based methods, generative adversarial networks (GANs), diffusion models, and Vision Transformer (ViT)-based approaches. A structured taxonomy is developed to classify existing methods into retrieval-based, generative-based, diffusion-based, transformer-based, and hybrid semantic mapping frameworks. A comparative analysis is performed based on accuracy, payload capacity, robustness, and visual quality. The analysis reveals a clear evolution from database-driven retrieval systems to advanced generative and transformer-based architectures. Recent diffusion and ViT-based models demonstrate superior robustness against compression, noise, and steganalysis attacks while maintaining high visual fidelity and improved semantic consistency. Deep semantic mapping significantly enhances the security and efficiency of coverless steganography. However, challenges remain in balancing payload capacity, computational complexity, and adversarial robustness. Future research should focus on hybrid architectures and real-time lightweight semantic encoding models.