ENHANCED IMAGE STEGANOGRAPHY WITH KERAS-STEGANOGAN: A TENSORFLOW-BASED GAN

Abstract

Image steganography is the process of embedding secret information within digital images such that the very presence of the message remains undetectable. Recent advances in deep learning, particularly in generative adversarial networks (GANs), have significantly improved both the payload capacity and perceptual quality of steganographic systems. The original SteganoGAN, implemented in Torch, achieved state-of-the-art performance by embedding up to 4.4 bits per pixel while maintaining strong resistance to steganalysis methods. However, the influence of the critic network on steganographic quality and learning stability remains insufficiently explored. This paper presents Keras-SteganoGAN, a TensorFlow-based reimplementation and extension of SteganoGAN, designed to systematically analyze the role of the critic in adversarial steganographic training. Two variants of the model–one incorporating a critic and one without–were trained and compared across three encoder architectures: basic convolutional, residual, and dense. Each configuration was trained over five epochs with message depths ranging from 1 to 6 bits, allowing a comprehensive study of trade-offs between payload capacity, image distortion, and decoding accuracy. Quantitative evaluation was conducted using standard image quality and steganographic metrics, including PSNR, SSIM, RS-BPP, and decoder accuracy. The results indicate that the inclusion of a critic improves perceptual quality and visual similarity at lower payloads, but its contribution diminishes as the message depth increases. These findings provide new insights into the interaction between encoder complexity, critic dynamics, and steganographic performance, offering guidance for the design of future GAN-based steganography systems.