StyleStego: a novel paradigm for high-capacity steganography using per layer noise maps in StyleGAN for cross-domain robustness

Steganography aims to embed secret information in digital media while preserving visual fidelity and resisting statistical detection. However, existing approaches typically face a trade-off between embedding capacity, imperceptibility, and robustness. This paper proposes StyleStego, a distribution-preserving generative steganography framework that exploits the stochastic per-layer noise mechanism of StyleGAN as an embedding medium. Secret bits are mapped to deterministic seeds that generate Gaussian noise patterns injected into multiple synthesis layers, enabling high-capacity embedding while maintaining natural image statistics. To improve robustness and generalization, the framework employs balanced multidomain training across five visual categories together with systematic augmentation. Experimental results show reliable extraction accuracy exceeding 90% at payloads up to 8 bits per pixel (bpp). Steganalysis evaluation demonstrates near-random detectability (PE ≈ 0.48–0.50) for payloads up to 4 bpp against SRM + EC and XuNet, while detectability decreases slightly at the maximum payload of 8 bpp. These results highlight the trade-off between maximum embedding capacity and secure operating payload while demonstrating that high-capacity generative steganography with strong perceptual fidelity and cross-domain robustness is achievable.