ISGI steganography scheme: image steganography scheme leveraging a novel cryptography algorithm and generative adversarial network

Abstract

Image steganography is a secure method for establishing covert communication. In this paper, we propose a novel image steganography scheme called ISGI, a novel image steganography scheme that combines chaos-based image encryption with deep learning autoencoder networks inspired by GAN architecture. This method first encrypts the secret image using a new chaos-based algorithm to enhance security, the IEPS encryption scheme, then encodes it into a cover image via a GAN-based autoencoder, where a generator creates realistic stego images and a discriminator distinguishes them from cover images. Adversarial training ensures the hidden message remains imperceptible. Evaluated on COCO and CelebA datasets, ISGI effectively hides one image inside another while maintaining high visual quality (PSNR 36–38 dB, SSIM 0.95–0.97), supports payload capacity over 23 bits per pixel, and achieves strong security metrics including high entropy ( 8), low correlation, and robust NPCR/UACI scores, demonstrating secure image embedding with minimal distortion.