GAN-Based SteRGAN Hides Secret Images Inside Cover Images With High Fidelity

Researchers have proposed SteRGAN, a generative adversarial network (GAN) based framework for image steganography, the practice of concealing data within cover images to enhance the confidentiality and protection of information. The work, published in Neural Computing and Applications, addresses long-standing weaknesses of traditional steganographic methods, which the authors note frequently suffer from low payload capacity, poor imperceptibility, or susceptibility to attacks.

The framework is built around two components, an Embedder module and an Extractor module, each of which contains its own generator and discriminator. This adversarial arrangement is central to the design: the networks are trained against one another, pushing the system to produce stego-images that look natural while still allowing hidden content to be recovered accurately.

A key contribution of the study is a deep network called RCBNet, proposed as the discriminator within the Embedder module. According to the authors, this architecture improves embedding capacity. RCBNet employs residual connections and dilated convolutions, two techniques intended to enhance feature extraction and preserve fine-grained visual details in the images the network processes.

These architectural choices, the paper reports, enable the generation of high-quality stego-images with minimal perceptual distortion. The model is trained with a composite loss function, an approach the authors say enhances the visual imperceptibility of the hidden content, meaning the stego image remains visually close to the original cover image.

Experimental evaluations were conducted on two widely used image datasets, ImageNet and Microsoft COCO. On ImageNet, SteRGAN achieved a Peak Signal-to-Noise Ratio (PSNR) of 42.50 dB and a Structural Similarity Index (SSIM) of 0.992 on the cover-stego image pair, and a PSNR of 42.50 dB with an SSIM of 0.989 on the secret-recovery image pair.

Results on Microsoft COCO were similar, though slightly lower on some measures. There, the framework yielded a PSNR of 42.34 dB and an SSIM of 0.979 on the cover-stego pair, and a PSNR of 41.99 dB with an SSIM of 0.991 on the secret-recovery pair. PSNR and SSIM are standard metrics for measuring how closely two images match, with higher values indicating greater similarity.

The authors state that these outcomes validate the effectiveness of the proposed framework compared with existing steganography methods. Because the same model performs both embedding and recovery at high fidelity across two different datasets, the results suggest the architecture can balance two competing goals that have traditionally been difficult to reconcile: hiding more data while leaving fewer visible traces.

The study is reported as an original article in Volume 38 of the journal. As with any newly proposed machine learning framework, the reported figures come from the authors’ own experimental evaluations on the two named datasets, and broader performance in real-world deployment would depend on factors beyond those benchmarks.

Subject of Research: A GAN-based framework for hiding secret images within cover images using a proposed discriminator network called RCBNet.

Article Title: SteRGAN: An Improvised Framework based on Generative Adversarial Network for Image Steganography

Article References: Sharma, Y., & Bhaumik, G. (2026). SteRGAN: An Improvised Framework based on Generative Adversarial Network for Image Steganography. Neural Computing and Applications, 38(20), Article 834. https://doi.org/10.1007/s00521-026-12559-1

Image Credits: AI Generated

DOI: 10.1007/s00521-026-12559-1

Keywords: image steganography, generative adversarial networks, SteRGAN, RCBNet, PSNR, SSIM, ImageNet, Microsoft COCO, deep learning, information security, neural networks, data hiding