Research map / Image generation / Text-to-image diffusion
SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images
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 SSHR, a diffusion-based generative image steganography method that replaces text prompts with reference images, uses an exact invertible reveal process, and derives reference-secret image related symmetric keys via an ECDHE-inspired exchange.
Problem
Diffusion-based generative image steganography suffers from stego images resembling secret images, degraded revealed secret images due to DDIM inversion inconsistency, and security risks from using text prompts as keys.
Approach
Replace text prompts with reference images as the guidance condition, make the nonlinear (Perona-Malik) diffusion reveal step exactly invertible, and generate per-pair symmetric keys from secret and reference image features via an ECDHE-style key exchange, which conditions convolution weights and modulation parameters.
Key results
- On DIV2K, ImageNet and COCO, SSHR reaches 48.56 dB, 49.52 dB and 47.67 dB PSNR on revealed secret images, beating the best cover-based baseline HiNet by 1.92–3.62 dB.
- On UniStega, revealed-image PSNR with the correct key is 47.04 dB versus 21.25 dB for CRoSS and 23.29 dB for DiffStega, while stego/secret PSNR drops to 8.96 dB.
- Steganalysis detection accuracy with SRNet and XuNet is about 51% for SSHR, lower than the generative baselines CRoSS and DiffStega (about 52–53%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use SSHR instead of HiNet | Generative image steganography · DIV2K | PSNR: 46.64 -> 48.56 | Helped |
| Use SSHR instead of HiNet | Generative image steganography · COCO | PSNR: 44.05 -> 47.67 | Helped |
| Use SSHR instead of HiNet | Generative image steganography · ImageNet | PSNR: 46.78 -> 49.52 | Helped |
| Use SSHR instead of DiffStega | Generative image steganography · UniStega, correct key | PSNR: 23.29 -> 47.04 | Helped |
| Use SSHR instead of DiffStega | Generative image steganography · UniStega, stego images | PSNR: 18.61 -> 8.96 | Helped |
| Use SSHR instead of DiffStega | Generative image steganography · UniStega | NIQE: 5.58 -> 5.46 | Helped |
| Use SSHR instead of DiffStega | Steganography model · Steganalysis, SRNet | detection accuracy (%): 52.03 -> 51.05 | Helped |
| Use SSHR instead of DiffStega | Steganography model · Steganalysis, XuNet | detection accuracy (%): 52.86 -> 50.93 | Helped |
| Use SSHR instead of DiffStega | Generative image steganography · UniStega, Gaussian blur | PSNR: 20.85 -> 24.01(3.16↑) | Helped |
| Use SSHR instead of DiffStega | Generative image steganography · UniStega, Gaussian noise | PSNR: 18.62 -> 23.95(4.98↑) | Helped |
| Use SSHR instead of DiffStega | Generative image steganography · UniStega, JPEG compression | PSNR: 21.16 -> 24.74(3.58↑) | Helped |
| Use SSHR instead of DiffStega | Generative image steganography · UniStega, Poisson noise | PSNR: 20.15 -> 23.96(3.81↑) | Helped |
Surprising findings
- The cover-based baseline HiNet and ISN are nearly as detectable as generative methods, but the generative baselines' stego images are no more natural than SSHR's by NIQE (5.46 vs 5.35 for HiNet), meaning NIQE differences are small.
Limitations
- Experiments are limited to 256×256 resolution.
- Robustness to distortions is only reported on the UniStega dataset with PSNR, and the NIQE gap versus baselines is small.
- Key exchange design is only briefly described; full protocol details are deferred to the supplementary material.
Experimental setup
Datasets: DIV2K COCO ImageNet UniStega
Benchmarks: DIV2K COCO ImageNet UniStega
Metrics: PSNR SSIM MAE RMSE NIQE detection accuracy AUC
Baselines: Baluja HiDDeN UDH Weng et al. ISN HiNet CRoSS DiffStega
Builds on: Perona-Malik diffusion HiNet CRoSS DiffStega ECDHE key exchange
Closest papers by meaning
- Provably Secure Image Robust Steganography via Cross-modal Error Correction (AAAI 2025 · similarity 0.89)
- StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography (AAAI 2024 · similarity 0.87)
- StegaStyleGAN: Towards Generic and Practical Generative Image Steganography (AAAI 2024 · similarity 0.86)
- MDDM: Practical Message-Driven Generative Image Steganography Based on Diffusion Models (ICML 2025 · similarity 0.85)
- Hiding Images in Diffusion Models by Editing Learned Score Functions (CVPR 2025 · similarity 0.85)
- Neural Cover Selection for Image Steganography (NeurIPS 2024 · similarity 0.85)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.