Comparative Evaluation of Deep Learning Encoder Architectures for Image-in-Image Steganography

Abstract

Abstract

With the increasing demand for secure and covert digital communication, image steganography has emerged as an effective technique for concealing sensitive information within digital images. Traditional steganographic methods based on spatial or transform domains often suffer from limited embedding capacity and vulnerability to image processing attacks. Recent advances in deep learning have enabled data-driven steganographic frameworks that jointly optimize information embedding and extraction while preserving visual quality. In this work, a unified deep learning–based image steganography framework is proposed to comparatively evaluate different convolutional neural network encoder architectures under identical experimental conditions. The framework follows an encoder–decoder paradigm in which a secret image is embedded into a cover image to generate a visually imperceptible stego image, and a common decoder reconstructs the hidden information. The encoder architectures are evaluated using objective metrics including Peak Signal-to-Noise Ratio, Mean Absolute Error, Visual Information Fidelity, and Normalized Cross-Correlation. Experimental results demonstrate that encoder design plays a critical role in determining steganographic performance, with densely connected architectures achieving superior imperceptibility and reconstruction accuracy. The study provides clear insights into architectural trade-offs and offers guidance for the design of efficient deep learning–based image steganography systems.