A two-stage depth-balanced GAN framework for secure and scalable high-capacity image steganography

Authors: Sultan, B., Wani, M.A.

Journal: Knowledge Based Systems

Publication Date: 09/10/2026

Volume: 351

ISSN: 0950-7051

DOI: 10.1016/j.knosys.2026.116605

Abstract:

Conventional and deep learning steganographic techniques typically involve encoding secret data within a cover image using a steganographic encoder, resulting in a stego image. This study introduces a novel two-stage approach for generating stego images, enhancing both capacity and security. The process comprises a cover reconstruction stage and a stego image construction stage, each utilizing architectures of varying depths. In the cover reconstruction stage, the input cover image is initially reconstructed to capture its distinctive features, adapting the size of the reconstructed cover according to the depth of this stage. This stage focuses on learning fundamental image features through a reconstruction loss.In the stego image construction stage, the reconstructed cover image and the secret message are combined to create the stego image. This stage employs three loss terms, including an adversarial loss, to incorporate the secret message. Importantly, the depth of these two stages exhibits an inverse relationship: a shallow architecture in the cover image reconstruction phase is balanced by a deep architecture in the stego image generation phase, and vice versa. This deliberate depth balancing ensures the stego image size does not exceed the cover image size, significantly enhancing embedding capacity.The unique aspect of this method is its focus on embedding data solely in the deeper layers of the Generator network, leaving the lower layers dedicated to image reconstruction. This approach boosts capacity without requiring complex architectures, thus reducing computational storage requirements.Experimental results on CelebA, COCO, and ImageNet datasets demonstrate that the proposed method achieves more than four times the embedding capacity of existing GAN-based approaches while preserving high imperceptibility (PSNR > 74 dB, SSIM > 0.95) and strong security (steganalyzer error ≈ 50%). Furthermore, the model ensures robust extraction accuracy at moderate payloads, confirming its practicality for real-world steganography applications.

Source: Scopus