Implementation And Comparative Analysis Of CNN, RNN, And GAN Models For Steganalysis In Color Images

Abstract

This paper examines three deep learning models—Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and Generative Adversarial Networks (GANs)—for the task of steganalysis in color image domains. Steganography, which involves embedding secret information into seemingly benign images, produces minute and highly localized irregularities that are difficult to identify using standard statistical or rule-based methods. To address these limitations, the study creates and implements domain-specific deep learning architectures that capture high-order spatial patterns, temporal embedding traces, and adversarial disparities induced by stego material. The models are trained and tested on the ALASKA2 dataset using a consistent preprocessing and augmentation workflow, and their detection accuracy, inference time, and model complexity are all evaluated. The findings demonstrate that CNN-based algorithms have the highest classification accuracy (99.52%), but GAN-based discriminators outperform in high-payload and adversarial situations. The findings highlight the significance of deep learning for effective steganalysis and lay the framework for future adversarial and ensemble-based forensic systems.