Research map / Image generation / Autoregressive and latent generation

StegaStyleGAN: Towards Generic and Practical Generative Image Steganography

AAAI 2024 · Image generation › Autoregressive and latent generation · approach: Robust diffusion watermarking · 32 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes StegaStyleGAN, a StyleGAN2-based generative image steganography scheme that modulates injected noise with secret bits via a distribution-preserving modulator, yielding provably secure lossless (Ls) and robust lossy (Ly) variants.

Problem

Generative image steganography is usually tailored to specific tasks and struggles to jointly achieve provable security, large embedding capacity, extraction accuracy, image quality, and robustness to channel attacks such as JPEG compression.

Approach

A Distribution-Preserving Secret Data Modulator (DP-SDM) fuses secret bits with StyleGAN2 injected noise as |N|·(2m−1), keeping the noise distribution unchanged so that KL divergence between cover and stego is zero; an Image Attack Simulator during training yields a robust variant, and GAN inversion gives conditional stego generation.

Key results

• StegaStyleGAN-Ls at 1 bpp on CelebA 128^2 reaches FID 6.12 and accuracy 97.75% with SRNet Pe=0.5, versus GSN's FID 13.29 and accuracy 97.53%.
• At 4 bpp on LSUN-bedroom 256^2, StegaStyleGAN-Ls achieves accuracy 95.65% and CRub 0.742 versus GSN's 72.13% and 0.146.
• StegaStyleGAN-Ly(32^2) under JPEG QF50 retains 99.36% accuracy at 0.0625 bpp, above CIS-Net's 95.46% at 0.031 bpp.

Measured effects reported in the paper

Surprising findings

• DP-SDM preserves security (Pe≈0.5) at a slight cost in accuracy and FID compared to the variant without it (Pe=1.0 without DP-SDM).

Limitations

• Accuracy drops notably at high capacity (e.g., 4 bpp at 128^2 gives 94.25%), requiring channel coding to reach reliable extraction.
• Security guarantee assumes secret bits are uniformly Bernoulli(0.5); performance is only checked under biased testing distributions.
• Lossy robustness results are reported with a single JPEG quality factor during training.

Experimental setup

Datasets: CelebA LSUN-bedroom

Benchmarks: CelebA LSUN-bedroom

Metrics: FID extraction accuracy steganalyzer detection error Pe code rate upper bound CRub

Baselines: GSN CIS-Net Xue and Wang 2021 Cao et al. 2020

Builds on: StyleGAN2 GSN CIS-Net LPIPS

Closest papers by meaning

• StegoGAN: Leveraging Steganography for Non-Bijective Image-to-Image Translation (CVPR 2024 · similarity 0.87)
• Provably Secure Image Robust Steganography via Cross-modal Error Correction (AAAI 2025 · similarity 0.87)
• SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images (ICML 2025 · similarity 0.86)
• MDDM: Practical Message-Driven Generative Image Steganography Based on Diffusion Models (ICML 2025 · similarity 0.85)
• StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography (AAAI 2024 · similarity 0.85)
• Robust Message Embedding via Attention Flow-Based Steganography (CVPR 2025 · similarity 0.83)