Research map / Image generation / Autoregressive and latent generation
Robust Message Embedding via Attention Flow-Based Steganography
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 RMSteg, which hides a QR code message in a host image using a normalizing flow with attention affine coupling blocks, invertible QR transition, and invertible token fusion, robust to printing/photography and other distortions.
Problem
Robust message-in-image steganography that embeds a QR code message into natural images with high stego quality, high capacity, and exact decoding after real-world distortions such as printing and photographing.
Approach
Integrates tokenized ViT image representations into a normalizing flow via attention affine coupling blocks (cross- and self-attention instead of CNN coupling), adds a learnable invertible token fusion matrix, and transforms the QR code conditioned on the host image before embedding.
Key results
- Ours achieves PSNR 32.883, SSIM 0.9109, LPIPS 0.0707 on stego images, versus ISN† at 32.175 / 0.8765 / 0.3266.
- Under printing and photography, RMSteg reaches TRA 1.000 with EMR 0.606, versus HiNet† TRA 0.970 and EMR 1.619.
- Maintains TRA above 0.8 for QR code v8 (688 bits) while keeping PSNR around 30.765.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add invertible QR Code transition (IQRT) (ablation) | RMSteg · Stego image and mixed distortions | LPIPS: 0.1059 -> 0.0707 | Helped |
| Add invertible QR Code transition (IQRT) (ablation) | RMSteg · Mixed distortions | TRA: 0.871 -> 0.856 | Hurt |
| Add invertible token fusion (ITF) (ablation) | RMSteg · Mixed distortions | LPIPS: 0.0919 -> 0.0707 | Helped |
| Add invertible token fusion (ITF) (ablation) | RMSteg · Mixed distortions | TRA: 0.833 -> 0.856 | Helped |
| Use tokenized representation and AttnFlow instead of CNN-based ISN (ablation) | RMSteg · Mixed distortions | LPIPS: 0.3209 -> 0.0707 | Helped |
| Use tokenized representation and AttnFlow instead of CNN-based HiNet (ablation) | RMSteg · Mixed distortions | TRA: 0.714 -> 0.856 | Helped |
| Increase the number of AACBs from 1 to 4 (ablation) | RMSteg · Mixed distortions | LPIPS: 0.1076 -> 0.0707 | Helped |
| Increase the number of AACBs from 1 to 4 (ablation) | RMSteg · Mixed distortions | TRA: 0.798 -> 0.856 | Helped |
| Use RMSteg instead of HiNet† | Robust message embedding · Mixed distortions | TRA: 0.677 -> 0.859 | Helped |
| Use RMSteg instead of ISN† | Robust message embedding · Printing | TRA: 0.960 -> 1.000 | Helped |
Surprising findings
- The model without IQRT slightly outperforms the full model in decoding accuracy under mixed distortions (TRA 0.871 vs 0.856), attributed to occasional module flips from the transition.
Limitations
- Decoding accuracy under some distortions (e.g., shooting distance 13 cm, mixed noise) is slightly below HiNet† in some settings.
- Robustness and capacity trade-off: higher QR versions reduce TRA and PSNR.
Experimental setup
Datasets: COCO train2017 COCO test2017 synthetic QR code images
Benchmarks: COCO test2017
Metrics: PSNR SSIM LPIPS TRA EMR
Baselines: ISN HiNet StegaStamp
Builds on: NICE GLOW HiNet ISN StegaStamp ArtCoder
Closest papers by meaning
- Provably Secure Image Robust Steganography via Cross-modal Error Correction (AAAI 2025 · similarity 0.86)
- MDDM: Practical Message-Driven Generative Image Steganography Based on Diffusion Models (ICML 2025 · similarity 0.85)
- SSHR: More Secure Generative Steganography with High-Quality Revealed Secret Images (ICML 2025 · similarity 0.84)
- Neural Cover Selection for Image Steganography (NeurIPS 2024 · similarity 0.84)
- StegoGAN: Leveraging Steganography for Non-Bijective Image-to-Image Translation (CVPR 2024 · similarity 0.84)
- StegFormer: Rebuilding the Glory of Autoencoder-Based Steganography (AAAI 2024 · similarity 0.84)
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.