Coverless Multi-Image Steganography Utilizing GANs & Diffusion Models

Abstract

Abstract This thesis proposes a novel framework for coverless multi-image steganography that leverages the complementary strengths of Generative Adversarial Networks and diffusion models. Traditional steganographic systems often depend on explicit cover modification, suffer from low capacity, or lack robustness against detection. Diffusion models excel in image synthesis but remain underexplored for multiimage steganography The proposed system successfully reconstructs two RGB secret images from generated container images with minimal perceptual loss, achieving PSNR values of up to 17 dB and SSIM scores of exceeding 0.6 even under standard Gaussian noise and JPEG compression. The hiding capacity reached 48 bits per pixel while maintaining resistance to steganalysis. This demonstrates the practical viability of the hybrid approach for secure, scalable, and robust coverless steganography. Limitations of prompt sensitivity, domain dependency, and recovery failure under high image similarity are identified, and future work explores flow-based conditioning to enhance recovery consistency. The findings demonstrate the viability of combining GAN and diffusion paradigms in scalable, multi-image, coverless steganographic systems.