StyleGAN2-Stego: Secure Coverless Image Steganography via Latent Space Encoding

الملخص

Coverless image steganography (CIS) enhances secrecy by avoiding direct modifications to existing images, unlike traditional methods that embed data by altering image content. However, many existing approaches struggle to balance payload capacity, visual quality, and resistance to steganalysis. This paper introduces StyleGAN2-Stego, a generative steganography framework that conceals secret messages in the latent space of a pre-trained StyleGAN2-ADA generator. The framework consists of an encoder that transforms a binary message into a latent perturbation vector, a fixed generator that synthesizes realistic stego images from the modified latent space, and a decoder that reconstructs the original message from the generated image. The encoder and decoder are trained jointly, while the generator remains frozen to preserve visual fidelity. Experimental results show that StyleGAN2-Stego achieves a payload of 0.5 bits per pixel, a message recovery accuracy of 98.72%, and strong resistance to detection by advanced steganalysis tools such as Xu-Net and Ye-Net. It also produces high-quality images, with a Fréchet Inception Distance (FID) of 6.78. These findings demonstrate that StyleGAN2-Stego effectively balances capacity, imperceptibility, and security, making it well-suited for secure image-based communication.