Deep Learning Techniques for Image Steganography: A Comprehensive Review

Abstract

Image steganography is an important aspect of secure communication that hides confidential messages withindigital images in a way that escapes detection. Conventional steganographic algorithms, including Least Significant Bit (LSB) and frequency domain-based approaches, have been found to have low embedding capacity, susceptibility to steganalysis attacks, and rigidity in terms of rule-based embedding processes. However, with the evolution of deep learning concepts, especially in the realms of convolutional neural networks (CNNs) and generative networks, image steganography has moved towards adaptive and data-driven models that have improved imperceptibility and robustness to a great extent. This review paper provides a detailed examination of the existing state-of-the-art deep learning-based models for image steganography, including CNN-based encoder-decoder models, GAN-based adaptive cost learning models, hybrid CNN-frequency domain models, and multi-layered steganographic models. The reviewed papers are critically compared in terms of embedding capacity, visual distortion, robustness to steganalysis attacks, computational complexity, and applicability. Based on this comparative analysis, the important research gaps in the existing models are identified. The review aims to act as a reference for researchers and students who would like to gain insight into the current developments in deep learning-based image steganography.In addition, the architectural trade-offs highlighted in this review provide practicalguidance for selecting suitable steganographic frameworks under different application constraints, including capacity, security, and deployment efficiency.